feat: re-platform agentic review onto executus + large-PR cost controls (#20)

Makes gadfly a consumer of executus (run.Executor compaction/bounding/budget/critic + fanout) and fixes the large-PR token burn in size-gated layers: paginated get_diff, downshift above GADFLY_HUGE_DIFF_BYTES, and a swarm-wide GADFLY_PR_BUDGET_SECS backstop. Small PRs untouched; advisory-only and the static binary preserved. Dogfood swarm reviewed it (6 models, 21 real findings graded + folded in).

Co-authored-by: Steve Dudenhoeffer <steve@stevedudenhoeffer.com>
Co-committed-by: Steve Dudenhoeffer <steve@stevedudenhoeffer.com>

.gitea/workflows/review-reusable.yml

@@ -60,6 +60,8 @@ on:
       worker_model:              { type: string, default: "" }   # GADFLY_WORKER_MODEL
       allowed_users:             { type: string, default: "" }   # GADFLY_ALLOWED_USERS (consumer-specific; set in your stub)
       trigger_phrase:            { type: string, default: "" }   # GADFLY_TRIGGER_PHRASE
+      consolidate:               { type: string, default: "" }   # GADFLY_CONSOLIDATE — "" => auto (one consensus comment for >=2 models); "0" => one comment per model
+      inline_review:             { type: string, default: "" }   # GADFLY_INLINE_REVIEW — "" => on (post a COMMENT-state PR review with inline comments on changed lines); "0" => off
       # Job wall-clock cap. 90 as a default: the 5-lens suite across a slow lane
       # (claude-code with extended thinking) over two passes can run long.
       timeout_minutes:           { type: number, default: 90 }
@@ -92,7 +94,7 @@ jobs:
     runs-on: ubuntu-latest
     timeout-minutes: ${{ inputs.timeout_minutes }}
     steps:
-      - uses: docker://gitea.stevedudenhoeffer.com/steve/gadfly:sha-c342bdb
+      - uses: docker://gitea.stevedudenhoeffer.com/steve/gadfly:sha-3095ebf
         env:
           # --- event context (from the CALLER's github.*) -------------------
           GITEA_API: ${{ github.server_url }}/api/v1/repos/${{ github.repository }}
@@ -143,3 +145,5 @@ jobs:
           GADFLY_WORKER_MODEL: ${{ inputs.worker_model }}
           GADFLY_ALLOWED_USERS: ${{ inputs.allowed_users }}
           GADFLY_TRIGGER_PHRASE: ${{ inputs.trigger_phrase }}
+          GADFLY_CONSOLIDATE: ${{ inputs.consolidate }}
+          GADFLY_INLINE_REVIEW: ${{ inputs.inline_review }}

CLAUDE.md

@@ -22,15 +22,21 @@ verifies each one against the actual code, and posts its findings as a comment.
 4. **Provider-agnostic.** Powered by [majordomo](https://gitea.stevedudenhoeffer.com/steve/majordomo),
    so it can target Ollama (local/cloud), OpenAI, Anthropic, Google, or any
    OpenAI/Ollama-compatible endpoint. Don't re-hardcode a single provider.
-5. **Portable & self-contained.** `cmd/gadfly` depends only on the Go stdlib + majordomo. Keep
+5. **Portable & self-contained.** `cmd/gadfly` depends only on the Go stdlib, majordomo, and
-   it that way — no heavyweight deps, no coupling to any one consumer repo (e.g. mort).
+   [executus](https://gitea.stevedudenhoeffer.com/steve/executus) (whose *core* — `run`/`compact`/
+   `model`/`fanout`/`tool` — is itself majordomo+stdlib only, so the binary stays static; do NOT
+   pull executus's `contrib/store` or any battery that drags in a DB driver). No heavyweight deps,
+   no coupling to any one consumer repo (e.g. mort). Gadfly is executus's canonical *light* consumer.
 
 ## Architecture
 
 ```
 cmd/gadfly/            the reviewer binary — pure producer of review markdown (stdout)
-  main.go              orchestration: loop specialists, each a review pass + adversarial recheck
+  main.go              orchestration: fan specialists out (executus/fanout), each a review pass + recheck
-  engine.go            reviewEngine abstraction: majordomo agent loop vs claude-code CLI shell-out
+  engine.go            reviewEngine abstraction: executus run.Executor (majordomo agent loop +
+                       compaction/bounding/budget/critic) vs claude-code CLI shell-out
+  executus.go          executus wiring: tool.Registry over the repo tools, the run.Executor build
+                       (compact + model context-limit threshold + per-PR budget + wrap-up critic)
   specialists.go       specialist lenses: built-ins, default suite, env + .gadfly.yml resolution
   auto.go              dynamic `auto` selection: a selector model picks lenses per-diff (may invent)
   delegate.go          worker-tier delegate_investigation tool (cheap sub-agent does legwork)
@@ -69,7 +75,7 @@ verdict. Verdict is one of: `No material issues found` / `Minor issues` / `Block
 ## Build / test
 
 ```sh
-go build ./cmd/gadfly      # needs read access to the private majordomo module
+go build ./cmd/gadfly      # needs read access to the private majordomo + executus modules
 go test ./...
 gofmt -l cmd/               # must be clean
 docker build -t gadfly:dev --secret id=REGISTRY_USER,env=REGISTRY_USER --secret id=REGISTRY_PASSWORD,env=REGISTRY_PASSWORD .
@@ -149,3 +155,21 @@ are actually exercised. OpenAI/Anthropic/Google come from majordomo's abstractio
   parallel, `cap` (from `GADFLY_PROVIDER_CONCURRENCY` else `GADFLY_CONCURRENCY`, default 1) bounds
   models-at-once within a lane. The review timeout (`GADFLY_TIMEOUT_SECS`) is **per-lens**, not
   shared across the suite — a slow model can't starve later lenses (the original timeout bug).
+- **Large-PR token burn**: the agent loop re-sends the whole transcript every step, so a giant
+  diff (the old `get_diff` dumped it untruncated, and it was embedded in both the review and
+  recheck task) was re-transmitted ~steps × lenses × passes × models times — a ~250 K-token PR
+  could drain a metered usage block in minutes. Fixed in three size-gated layers (small PRs
+  untouched): paginated `get_diff` + `executus/compact` compaction in the binary; an
+  `entrypoint.sh` downshift above `GADFLY_HUGE_DIFF_BYTES` (one cheap model, fewer lenses/steps,
+  no recheck); and a swarm-wide `GADFLY_PR_BUDGET_SECS` wall-clock backstop. Compaction's threshold
+  is intentionally LOW (`GADFLY_COMPACT_RATIO` 0.45, not executus's 0.7) because the burning
+  transcript on the embedded path rarely reaches 0.7×context.
+- **executus re-platform**: the in-process review path runs through `executus/run`'s `run.Executor`
+  (compaction, run-bounding, `Ports.Budget`, the wrap-up nudge as `Ports.Critic`), wiring it in
+  `cmd/gadfly/executus.go`. Gadfly KEEPS its own `model.go` resolution (so `GADFLY_ENDPOINT_<NAME>`
+  http aliases + the claude-code engine survive) and only hands `run.Executor` the already-resolved
+  model via a trivial resolver — do NOT route review-model resolution through
+  `model.ParseModelForContext` (it bypasses gadfly's endpoint aliases). `run.Result` exposes no
+  transcript, so the old transcript-based forced-finalization fallback is gone; the wrap-up critic
+  nudge is the remaining "always emit something" mechanism. The claude-code engine still shells out
+  and is unaffected.

Dockerfile

@@ -24,7 +24,10 @@ RUN --mount=type=cache,target=/go/pkg/mod \
     go build -trimpath -ldflags="-s -w" -o /out/gadfly ./cmd/gadfly
 
 FROM alpine:3.20
-RUN apk add --no-cache bash git curl jq ca-certificates nodejs npm
+# procps provides pkill/pgrep, which entrypoint.sh's per-PR wall-clock backstop
+# (GADFLY_PR_BUDGET_SECS) uses to stop the review subtrees — busybox's applets
+# are not guaranteed to include them.
+RUN apk add --no-cache bash git curl jq ca-certificates nodejs npm procps
 # Bundle the Claude Code CLI so the `claude-code` review engine works out of the
 # box (GADFLY_MODELS=claude-code or claude-code/<model>). This adds Node + the
 # CLI to the image (notably larger); ollama-only users pay the size but nothing

README.md

@@ -277,6 +277,43 @@ every `GADFLY_STATUS_POLL_SECS` (default 12s) until the swarm finishes. It's adv
 best-effort — the per-model findings comments are unaffected — and entirely separate from those.
 Turn it off with `GADFLY_STATUS_BOARD=0`.
 
+### Consensus consolidation
+
+With **two or more models**, posting one comment each means a reader faces N walls of prose that
+mostly agree. Instead Gadfly consolidates: every model writes its findings to a shared file, and
+after the whole swarm finishes a single pass clusters those findings by location, counts **how
+many models independently flagged each one**, and posts **one consensus comment**:
+
+```
+## 🪰 Gadfly review — consensus across 7 models
+
+**Verdict: Blocking issues found** · 9 findings (3 with multi-model agreement)
+
+| | Finding | Where | Models | Lens |
+|--|--|--|--|--|
+| 🔴 | Auth bypass: token not verified | `auth/login.go:42` | 6/7 | security |
+| 🟠 | Unbounded retry loop | `sync/worker.go:88` | 3/7 | error-handling |
+
+<details><summary>4 single-model findings (lower confidence)</summary> … </details>
+<details><summary>Per-model detail</summary> … each model's full review, folded … </details>
+```
+
+Cross-model agreement is the strongest real-vs-false-positive signal available, so findings are
+ranked by it (a lone low-severity finding folds away; a lone *critical* still surfaces). The
+per-model comments are suppressed in this mode — each model's full review is preserved, folded,
+inside the consensus comment — and nothing is lost: if consolidation can't run, Gadfly falls
+back to posting the per-model comments. Controlled by `GADFLY_CONSOLIDATE`: `auto` (default — on
+for ≥2 models), `1` (force on), `0` (force off, one comment per model). Single-model runs are
+unaffected.
+
+**Inline PR review.** Alongside the consensus comment, Gadfly also posts a single Gitea **pull
+review** (state `COMMENT` — advisory, **never** request-changes or approve, so it can't block a
+merge) whose inline comments anchor each consensus finding to the exact changed line it's about.
+Only findings that land on a line in the diff are anchored (Gitea rejects comments off the diff);
+the rest stay in the consensus comment. A re-run replaces the previous review instead of stacking.
+It's the "reviewer integrated with Gitea" without the blocking — turn it off with
+`GADFLY_INLINE_REVIEW=0`.
+
 ### Triggers
 
 1. A **new/reopened/ready** non-draft PR — automatic.
@@ -359,9 +396,20 @@ The reviewer binary reads these (the stub/entrypoint set sane defaults):
 | `GADFLY_TIMEOUT_SECS` | 300 | deadline **per specialist lens** (review+recheck) |
 | `GADFLY_RECHECK` | on | set `0`/`false` to skip the recheck pass |
 | `GADFLY_RECHECK_MAX_STEPS` | 16 | recheck-pass step cap |
-| `GADFLY_MAX_DIFF_CHARS` | 60000 | diff chars embedded in the prompt (full diff via `get_diff`) |
+| `GADFLY_MAX_DIFF_CHARS` | 60000 | diff chars embedded in the **review** prompt (the full diff is reachable via the paginated `get_diff` tool, scoped per file with its `path` arg) |
+| `GADFLY_RECHECK_DIFF_CHARS` | 20000 | diff chars embedded in the **recheck** prompt (smaller — the recheck pages `get_diff` for the hunks it verifies) |
+| `GADFLY_COMPACT` | on | context compaction (via [executus](https://gitea.stevedudenhoeffer.com/steve/executus)): fold the transcript's runaway middle into a summary as it nears the model's context window, so a big diff + accumulating tool output can't balloon every step. `0` disables |
+| `GADFLY_COMPACT_RATIO` | 0.45 | fraction of the model's context window at which compaction fires |
+| `GADFLY_COMPACT_MODEL` | worker, else review model | cheap model the compactor uses to summarize the folded middle |
+| `GADFLY_COMPACT_KEEP_RECENT` | 8 | most-recent messages kept verbatim during compaction |
+| `GADFLY_COMPACT_SUMMARY_WORDS` | 200 | word cap on the compaction summary |
+| `GADFLY_MODEL_CONTEXT_TOKENS` | *(auto)* | override the model's context-window size (tokens) for the compaction threshold; set it for self-hosted endpoints executus can't introspect (Ollama Cloud models resolve automatically) |
+| `GADFLY_PR_TOKEN_BUDGET` | — | per-model token ceiling for this PR; once spent, remaining lenses/passes are skipped (advisory). 0 = off |
+| `GADFLY_PR_TIME_BUDGET_SECS` | — | per-model wall-clock ceiling for this PR (advisory). 0 = off |
 | `GADFLY_STATUS_BOARD` | on | set `0` to disable the live status-board comment |
 | `GADFLY_STATUS_POLL_SECS` | 12 | how often the status board re-renders/upserts |
+| `GADFLY_CONSOLIDATE` | `auto` | cross-model consensus comment: `auto` (on for ≥2 models), `1` (force on), `0` (off — one comment per model) |
+| `GADFLY_INLINE_REVIEW` | on | when consolidating, also post a `COMMENT`-state PR review with inline comments on changed lines; `0` disables |
 | `GADFLY_TRIGGER_PHRASE` | `@gadfly review` | comment phrase that re-triggers |
 | `GADFLY_ALLOWED_USERS` | *(collaborators)* | comma-separated allow-list for comment triggers |
 | `GADFLY_FINDINGS_URL` | — | gadfly-reports store base URL; set to enable findings telemetry (off when empty) |
@@ -369,6 +417,37 @@ The reviewer binary reads these (the stub/entrypoint set sane defaults):
 | `GADFLY_REPO` | *(from `GITEA_API`)* | `owner/repo` slug stamped on emitted runs/findings (set by `entrypoint.sh`) |
 | `GADFLY_PR` | *(from event)* | PR number stamped on emitted runs/findings (set by `entrypoint.sh`) |
 
+### Large-PR cost controls
+
+A very large diff is the one thing that can blow the budget: every review step
+re-sends it, multiplied across models × lenses × passes × steps (a single
+~250 K-token PR can otherwise burn a whole metered usage block). Gadfly handles
+big PRs in three layers, all **size-gated so small PRs are untouched**:
+
+1. **Paginated `get_diff` + compaction** (reviewer binary, on by default) —
+   `get_diff` returns a paginated, optionally per-file window instead of the whole
+   diff, and once a transcript nears the model's context window its middle is
+   folded into a summary (powered by [executus](https://gitea.stevedudenhoeffer.com/steve/executus)'s
+   `compact`). Tune with the `GADFLY_COMPACT_*` knobs above.
+2. **Downshift** (`entrypoint.sh`) — above `GADFLY_HUGE_DIFF_BYTES` the whole fleet
+   collapses to a single cheap model + a focused lens subset, fewer steps, and no
+   recheck. A finished shallow review beats a budget-nuking one, and the posted
+   comment says so.
+3. **Hard backstop** (`entrypoint.sh`) — `GADFLY_PR_BUDGET_SECS` is a wall-clock
+   ceiling across the *entire* fleet; on expiry the review is stopped and whatever
+   was found so far is posted. Like everything else, it never fails CI.
+
+| Env | Default | Meaning |
+|-----|---------|---------|
+| `GADFLY_HUGE_DIFF_BYTES` | 600000 | downshift the fleet when the PR diff exceeds this many bytes (0 = never downshift) |
+| `GADFLY_HUGE_DIFF_MODELS` | first model | model(s) to run on a downshifted huge PR |
+| `GADFLY_HUGE_DIFF_SPECIALISTS` | `security,correctness,error-handling` | lenses on a downshifted huge PR |
+| `GADFLY_HUGE_DIFF_MAX_STEPS` | 12 | review step cap on a huge PR |
+| `GADFLY_HUGE_DIFF_RECHECK_MAX_STEPS` | 8 | recheck step cap on a huge PR |
+| `GADFLY_HUGE_DIFF_RECHECK` | 0 | run the recheck pass on a huge PR (off by default) |
+| `GADFLY_HUGE_DIFF_MAX_DIFF_CHARS` | 20000 | embedded review-diff chars on a huge PR |
+| `GADFLY_PR_BUDGET_SECS` | — | swarm-wide wall-clock backstop; stops the whole fleet when reached (0 = off) |
+
 ## Findings telemetry (optional)
 
 Gadfly can record what it found so model quality can be tracked over time. It is
@@ -392,7 +471,7 @@ code.
 ## Building locally
 
 ```sh
-go build ./cmd/gadfly      # needs read access to the private majordomo module
+go build ./cmd/gadfly      # needs read access to the private majordomo + executus modules
 go test ./...
 ```
 

cmd/gadfly/consensus.go

@@ -0,0 +1,477 @@
+package main
+
+// Cross-model consensus consolidation. The swarm runs each model independently
+// (entrypoint.sh fans them out across provider lanes); historically each model
+// posted its OWN comment, so a reader faced N walls of prose that mostly agreed.
+//
+// Instead, every model writes its findings to a shared directory
+// (GADFLY_FINDINGS_OUT, one JSON file per model), and after the whole swarm
+// finishes a single consolidation pass (GADFLY_CONSOLIDATE_DIR) clusters those
+// findings by location, counts how many models independently flagged each one,
+// and renders ONE comment: an agreement-ranked table up top (cross-model
+// agreement is the strongest real-vs-false-positive signal we have), with each
+// model's full review folded below for drill-down.
+//
+// This file owns: the per-model artifact (modelFindings), writing it
+// (writeFindingsOut), reading the directory back, clustering, and rendering the
+// consensus markdown (renderConsensus). It depends only on the stdlib.
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+)
+
+// consensusMarker tags the single consolidated comment so entrypoint.sh can
+// upsert it in place across re-runs (mirrors run.sh's per-model marker).
+const consensusMarker = "<!-- gadfly-consensus -->"
+
+// modelFindings is the per-model artifact written to GADFLY_FINDINGS_OUT. It
+// carries enough to rebuild the consolidated comment without re-running models:
+// the structured findings (for clustering) plus the full rendered review (for
+// the folded per-model drill-down).
+type modelFindings struct {
+	Model    string       `json:"model"`
+	Provider string       `json:"provider"`
+	Verdict  string       `json:"verdict"`  // worst lens verdict, as a label
+	Errored  bool         `json:"errored"`  // the model produced no usable review (every lens failed, or the run crashed)
+	Markdown string       `json:"markdown"` // full rendered per-model review (findings block already stripped)
+	Findings []outFinding `json:"findings"`
+}
+
+// outFinding is one finding in the per-model artifact (a flattened `finding`
+// plus its lens).
+type outFinding struct {
+	Lens       string `json:"lens"`
+	File       string `json:"file"`
+	Line       int    `json:"line"`
+	Severity   string `json:"severity"`
+	Confidence string `json:"confidence"`
+	Title      string `json:"title"`
+	Detail     string `json:"detail"`
+}
+
+// collectFindings returns a lens result's findings with severity always filled
+// in: per-finding when the structured block supplied it, else derived from the
+// lens verdict (so heuristic-scraped findings still carry a canonical word). A
+// clean or errored lens yields nothing. Shared by the telemetry emit and the
+// per-model findings file so both agree on what a finding is.
+func collectFindings(r specialistResult) []finding {
+	if r.errored || r.verdict == verdictClean {
+		return nil
+	}
+	fs := extractStructuredFindingsOrScrape(r)
+	lensSev := r.verdict.severity()
+	for i := range fs {
+		if fs[i].severity == "" {
+			fs[i].severity = lensSev
+		}
+	}
+	return fs
+}
+
+// writeFindingsOut writes this model's findings + rendered review to
+// GADFLY_FINDINGS_OUT for the later consolidation pass. No-op unless the env is
+// set. Best-effort: any error is logged to stderr and never affects the review
+// (it runs after the markdown is already on stdout).
+func writeFindingsOut(results []specialistResult) {
+	path := strings.TrimSpace(os.Getenv("GADFLY_FINDINGS_OUT"))
+	if path == "" {
+		return
+	}
+	mf := modelFindings{
+		Model:    strings.TrimSpace(os.Getenv("GADFLY_MODEL")),
+		Provider: modelProvider(),
+		Verdict:  worstVerdict(results).label(),
+		Errored:  allErrored(results),
+		Markdown: renderConsolidated(results),
+	}
+	for _, r := range results {
+		for _, f := range collectFindings(r) {
+			mf.Findings = append(mf.Findings, outFinding{
+				Lens:       r.spec.Name,
+				File:       f.file,
+				Line:       f.line,
+				Severity:   f.severity,
+				Confidence: f.confidence,
+				Title:      f.title,
+				Detail:     f.detail,
+			})
+		}
+	}
+	data, err := json.Marshal(mf)
+	if err != nil {
+		fmt.Fprintln(os.Stderr, "gadfly: marshal findings out:", err)
+		return
+	}
+	// Defensive: make sure the parent dir exists (entrypoint creates it, but a
+	// missing dir would otherwise silently drop this model from the consensus).
+	if dir := filepath.Dir(path); dir != "" {
+		_ = os.MkdirAll(dir, 0o755)
+	}
+	if err := os.WriteFile(path, data, 0o644); err != nil {
+		fmt.Fprintln(os.Stderr, "gadfly: write findings out:", err)
+	}
+}
+
+// allErrored reports whether every lens of a review failed (so the model
+// produced no usable findings). Such a model is recorded but excluded from the
+// consensus agreement denominator — counting it would dilute every ratio with a
+// model that never actually reviewed.
+func allErrored(results []specialistResult) bool {
+	if len(results) == 0 {
+		return true
+	}
+	for _, r := range results {
+		if !r.errored {
+			return false
+		}
+	}
+	return true
+}
+
+// runConsolidate is the consolidation entry point (GADFLY_CONSOLIDATE_DIR set):
+// read every per-model artifact in the directory, render the consensus comment
+// to stdout. Errors are fatal to THIS process only — entrypoint.sh treats a
+// failed consolidation as advisory and falls back to per-model comments.
+func runConsolidate() error {
+	dir := strings.TrimSpace(os.Getenv("GADFLY_CONSOLIDATE_DIR"))
+	if dir == "" {
+		return errors.New("GADFLY_CONSOLIDATE_DIR is empty")
+	}
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		return fmt.Errorf("read consolidate dir: %w", err)
+	}
+	var models []modelFindings
+	for _, e := range entries {
+		if e.IsDir() || !strings.HasSuffix(e.Name(), ".json") {
+			continue
+		}
+		data, err := os.ReadFile(filepath.Join(dir, e.Name()))
+		if err != nil {
+			fmt.Fprintln(os.Stderr, "gadfly: read", e.Name(), err)
+			continue
+		}
+		var mf modelFindings
+		if err := json.Unmarshal(data, &mf); err != nil {
+			fmt.Fprintln(os.Stderr, "gadfly: parse", e.Name(), err)
+			continue
+		}
+		if strings.TrimSpace(mf.Model) == "" {
+			continue
+		}
+		models = append(models, mf)
+	}
+	if len(models) == 0 {
+		return errors.New("no model findings to consolidate")
+	}
+	// Cluster once, then render the consensus comment and (best-effort) the inline
+	// PR review from the same clusters so the two views can't drift.
+	clusters := clusterFindings(models)
+
+	// Lead with the marker so entrypoint.sh can upsert this comment in place
+	// (same pattern as run.sh's per-model marker); it appends the advisory footer.
+	fmt.Println(consensusMarker)
+	fmt.Println(renderConsensus(models, clusters))
+
+	// Inline PR review (COMMENT state) anchoring findings to changed lines.
+	// Best-effort: no-op without diff/API creds, never affects stdout/exit.
+	postInlineReview(clusters)
+	return nil
+}
+
+// cluster is a group of findings (across models) judged to be the same issue:
+// same file, lines within lineTolerance of the cluster's current span. The span
+// [line,maxLine] slides as members join, so a chain of nearby findings merges
+// instead of splitting once it drifts past the first line.
+type cluster struct {
+	file     string
+	line     int // representative (smallest) line
+	maxLine  int // largest line in the cluster — the span's upper edge
+	severity string
+	title    string
+	detail   string // highest-severity report's detail; rendered in the inline review comment
+	models   map[string]bool
+	lenses   map[string]bool
+}
+
+// findingRef is one model's finding (carrying which model reported it), used
+// while grouping findings into clusters.
+type findingRef struct {
+	f     outFinding
+	model string
+}
+
+// lineTolerance: a finding in the same file within this many lines of a
+// cluster's current span is treated as the same issue (models often cite a line
+// or two apart).
+const lineTolerance = 3
+
+// renderConsensus builds the single consolidated comment body from every model's
+// findings. It does NOT emit the marker or advisory footer — entrypoint.sh wraps
+// it (mirroring run.sh's per-model framing).
+func renderConsensus(models []modelFindings, clusters []cluster) string {
+	// effective = models that actually produced a review. Errored models are
+	// shown (folded, below) but excluded from the agreement denominator so a
+	// failed model doesn't dilute every ratio.
+	effective := 0
+	for _, m := range models {
+		if !m.Errored {
+			effective++
+		}
+	}
+	errored := len(models) - effective
+
+	worst := verdictClean
+	for _, m := range models {
+		if v := parseVerdict(m.Verdict); v > worst {
+			worst = v
+		}
+	}
+
+	// Partition in one pass: "headline" findings (multi-model agreement, OR a
+	// lone CRITICAL) vs folded "single-model" lower-confidence findings. Also
+	// count multi-model agreements for the summary line.
+	var headline, folded []cluster
+	agreed := 0
+	for _, c := range clusters {
+		if len(c.models) >= 2 {
+			agreed++
+		}
+		if len(c.models) >= 2 || sevRank(c.severity) >= sevRank("critical") {
+			headline = append(headline, c)
+		} else {
+			folded = append(folded, c)
+		}
+	}
+
+	var b strings.Builder
+	fmt.Fprintf(&b, "## 🪰 Gadfly review — consensus across %d model%s", effective, plural(effective))
+	if errored > 0 {
+		fmt.Fprintf(&b, " (%d failed)", errored)
+	}
+	b.WriteString("\n\n")
+	fmt.Fprintf(&b, "**Verdict: %s** · %d finding%s (%d with multi-model agreement)\n",
+		worst.label(), len(clusters), plural(len(clusters)), agreed)
+
+	if len(headline) > 0 {
+		b.WriteString("\n| | Finding | Where | Models | Lens |\n|--|--|--|--|--|\n")
+		for _, c := range headline {
+			fmt.Fprintf(&b, "| %s | %s | `%s` | %d/%d | %s |\n",
+				sevIcon(c.severity), mdCell(c.title), mdCell(location(c.file, c.line)),
+				len(c.models), effective, mdCell(lensList(c.lenses)))
+		}
+	} else if len(clusters) == 0 {
+		b.WriteString("\nNo material issues found by consensus.\n")
+	}
+	// else: only single-model findings — they're shown folded below, so don't
+	// claim "no material issues" (there are some, just none with consensus).
+
+	if len(folded) > 0 {
+		fmt.Fprintf(&b, "\n<details><summary>%d single-model finding%s (lower confidence)</summary>\n\n",
+			len(folded), plural(len(folded)))
+		b.WriteString("| | Finding | Where | Model | Lens |\n|--|--|--|--|--|\n")
+		for _, c := range folded {
+			fmt.Fprintf(&b, "| %s | %s | `%s` | %s | %s |\n",
+				sevIcon(c.severity), mdCell(c.title), mdCell(location(c.file, c.line)),
+				mdCell(oneModel(c.models)), mdCell(lensList(c.lenses)))
+		}
+		b.WriteString("\n</details>\n")
+	}
+
+	// Per-model full reviews, folded for drill-down (nothing is lost).
+	b.WriteString("\n<details><summary>Per-model detail</summary>\n")
+	for _, m := range models {
+		body := strings.TrimSpace(m.Markdown)
+		if body == "" {
+			body = "_(no output)_"
+		}
+		verdict := m.Verdict
+		if m.Errored {
+			verdict = "⚠️ reviewer failed"
+		}
+		fmt.Fprintf(&b, "\n<details><summary><b>%s</b> (%s) — %s</summary>\n\n%s\n\n</details>\n",
+			mdCell(m.Model), mdCell(m.Provider), verdict, body)
+	}
+	b.WriteString("\n</details>")
+	return b.String()
+}
+
+// clusterFindings groups every model's findings into cross-model clusters,
+// sorted by agreement (desc), then severity (desc), then location.
+func clusterFindings(models []modelFindings) []cluster {
+	// Group by file, then greedily merge by line proximity.
+	byFile := map[string][]findingRef{}
+	for _, m := range models {
+		for _, f := range m.Findings {
+			if strings.TrimSpace(f.File) == "" {
+				continue
+			}
+			byFile[f.File] = append(byFile[f.File], findingRef{f, m.Model})
+		}
+	}
+
+	var clusters []cluster
+	for file, items := range byFile {
+		sort.SliceStable(items, func(i, j int) bool { return items[i].f.Line < items[j].f.Line })
+		// Cluster within THIS file only (clusters never span files), so the inner
+		// scan is over same-file clusters, not every cluster seen so far.
+		var fileClusters []cluster
+		for _, it := range items {
+			placed := false
+			for ci := range fileClusters {
+				c := &fileClusters[ci]
+				// Join if the line falls within the cluster's span, widened by the
+				// tolerance on both edges — so the window slides as the span grows.
+				if it.f.Line >= c.line-lineTolerance && it.f.Line <= c.maxLine+lineTolerance {
+					mergeIntoCluster(c, it.f, it.model)
+					placed = true
+					break
+				}
+			}
+			if !placed {
+				c := cluster{
+					file:     file,
+					line:     it.f.Line,
+					maxLine:  it.f.Line,
+					severity: it.f.Severity,
+					title:    it.f.Title,
+					detail:   it.f.Detail,
+					models:   map[string]bool{},
+					lenses:   map[string]bool{},
+				}
+				mergeIntoCluster(&c, it.f, it.model)
+				fileClusters = append(fileClusters, c)
+			}
+		}
+		clusters = append(clusters, fileClusters...)
+	}
+
+	sort.SliceStable(clusters, func(i, j int) bool {
+		if len(clusters[i].models) != len(clusters[j].models) {
+			return len(clusters[i].models) > len(clusters[j].models)
+		}
+		if sevRank(clusters[i].severity) != sevRank(clusters[j].severity) {
+			return sevRank(clusters[i].severity) > sevRank(clusters[j].severity)
+		}
+		if clusters[i].file != clusters[j].file {
+			return clusters[i].file < clusters[j].file
+		}
+		return clusters[i].line < clusters[j].line
+	})
+	return clusters
+}
+
+// mergeIntoCluster folds one finding into a cluster: union the model/lens sets,
+// widen the [line,maxLine] span, and keep the highest-severity report's title.
+func mergeIntoCluster(c *cluster, f outFinding, model string) {
+	if model != "" {
+		c.models[model] = true
+	}
+	if f.Lens != "" {
+		c.lenses[f.Lens] = true
+	}
+	if f.Line > 0 && (c.line == 0 || f.Line < c.line) {
+		c.line = f.Line
+	}
+	if f.Line > c.maxLine {
+		c.maxLine = f.Line
+	}
+	// Backfill an empty title/detail from any report, regardless of severity, so a
+	// higher-severity-but-terse finding doesn't leave the cluster without context.
+	if strings.TrimSpace(c.title) == "" && strings.TrimSpace(f.Title) != "" {
+		c.title = f.Title
+	}
+	if strings.TrimSpace(c.detail) == "" && strings.TrimSpace(f.Detail) != "" {
+		c.detail = f.Detail
+	}
+	// A strictly-higher-severity report takes over the title/detail.
+	if sevRank(f.Severity) > sevRank(c.severity) {
+		c.severity = f.Severity
+		if strings.TrimSpace(f.Title) != "" {
+			c.title = f.Title
+		}
+		if strings.TrimSpace(f.Detail) != "" {
+			c.detail = f.Detail
+		}
+	}
+}
+
+// sevRank orders the canonical severity words for sorting/comparison.
+func sevRank(s string) int {
+	switch strings.ToLower(strings.TrimSpace(s)) {
+	case "critical":
+		return 5
+	case "high":
+		return 4
+	case "medium":
+		return 3
+	case "small":
+		return 2
+	case "trivial":
+		return 1
+	default:
+		return 0
+	}
+}
+
+// sevIcon is the at-a-glance severity badge for the consensus table.
+func sevIcon(s string) string {
+	switch strings.ToLower(strings.TrimSpace(s)) {
+	case "critical", "high":
+		return "🔴"
+	case "medium":
+		return "🟠"
+	case "small":
+		return "🟡"
+	default:
+		return "⚪"
+	}
+}
+
+func location(file string, line int) string {
+	if line > 0 {
+		return fmt.Sprintf("%s:%d", file, line)
+	}
+	return file
+}
+
+func lensList(lenses map[string]bool) string {
+	out := make([]string, 0, len(lenses))
+	for l := range lenses {
+		out = append(out, l)
+	}
+	sort.Strings(out)
+	return strings.Join(out, ", ")
+}
+
+func oneModel(models map[string]bool) string {
+	for m := range models {
+		return m
+	}
+	return ""
+}
+
+// mdCell makes a string safe for a one-line markdown table cell: collapse
+// newlines, escape pipes (which delimit columns), and neutralize backticks
+// (a stray one would break an inline-code span — a backslash can't escape it
+// inside code, so replace with an apostrophe). Inputs are model-influenced, so
+// this keeps a malformed file path or title from breaking the table.
+func mdCell(s string) string {
+	s = strings.ReplaceAll(s, "\n", " ")
+	s = strings.ReplaceAll(s, "|", "\\|")
+	s = strings.ReplaceAll(s, "`", "'")
+	return strings.TrimSpace(s)
+}
+
+func plural(n int) string {
+	if n == 1 {
+		return ""
+	}
+	return "s"
+}

cmd/gadfly/consensus_test.go

@@ -0,0 +1,217 @@
+package main
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+func TestClusterFindingsAgreementAndTolerance(t *testing.T) {
+	models := []modelFindings{
+		{Model: "m1", Verdict: "Blocking issues found", Findings: []outFinding{
+			{Lens: "security", File: "a.go", Line: 10, Severity: "high", Title: "auth bypass"},
+			{Lens: "perf", File: "b.go", Line: 5, Severity: "trivial", Title: "tiny nit"},
+		}},
+		{Model: "m2", Verdict: "Minor issues", Findings: []outFinding{
+			{Lens: "security", File: "a.go", Line: 11, Severity: "critical", Title: "auth bypass (crit)"}, // within tolerance of a.go:10
+		}},
+		{Model: "m3", Verdict: "Minor issues", Findings: []outFinding{
+			{Lens: "correctness", File: "a.go", Line: 10, Severity: "medium", Title: "auth bypass"},
+		}},
+	}
+	clusters := clusterFindings(models)
+	if len(clusters) != 2 {
+		t.Fatalf("want 2 clusters (a.go:10±, b.go:5), got %d: %+v", len(clusters), clusters)
+	}
+	// First cluster (highest agreement) is the a.go auth one: 3 models, severity
+	// escalated to critical, representative line the smallest (10).
+	c := clusters[0]
+	if len(c.models) != 3 {
+		t.Errorf("want 3 models on the top cluster, got %d", len(c.models))
+	}
+	if c.severity != "critical" {
+		t.Errorf("want escalated severity critical, got %q", c.severity)
+	}
+	if c.line != 10 {
+		t.Errorf("want representative line 10, got %d", c.line)
+	}
+	if !c.lenses["security"] || !c.lenses["correctness"] {
+		t.Errorf("want union of lenses, got %v", c.lenses)
+	}
+}
+
+func TestRenderConsensusFoldsSingleModelNits(t *testing.T) {
+	models := []modelFindings{
+		{Model: "m1", Provider: "p", Verdict: "Blocking issues found", Markdown: "m1 detail", Findings: []outFinding{
+			{Lens: "security", File: "a.go", Line: 10, Severity: "high", Title: "auth bypass"},
+			{Lens: "perf", File: "b.go", Line: 5, Severity: "trivial", Title: "tiny nit"},
+		}},
+		{Model: "m2", Provider: "p", Verdict: "Minor issues", Markdown: "m2 detail", Findings: []outFinding{
+			{Lens: "security", File: "a.go", Line: 10, Severity: "high", Title: "auth bypass"},
+		}},
+	}
+	out := renderConsensus(models, clusterFindings(models))
+	// Headline table: the agreed finding with a 2/2 badge.
+	if !strings.Contains(out, "2/2") {
+		t.Errorf("expected a 2/2 agreement badge in headline:\n%s", out)
+	}
+	if !strings.Contains(out, "auth bypass") || !strings.Contains(out, "a.go:10") {
+		t.Errorf("headline missing the consensus finding:\n%s", out)
+	}
+	// The lone trivial finding is folded, not in the headline table.
+	if !strings.Contains(out, "single-model finding") {
+		t.Errorf("expected a folded single-model section:\n%s", out)
+	}
+	// Per-model detail is preserved (folded).
+	if !strings.Contains(out, "m1 detail") || !strings.Contains(out, "m2 detail") {
+		t.Errorf("per-model detail not preserved:\n%s", out)
+	}
+}
+
+func TestRenderConsensusHighSeverityLoneFindingStaysHeadline(t *testing.T) {
+	// A single model, single critical finding must still surface in the headline
+	// (not be folded as "low confidence").
+	models := []modelFindings{
+		{Model: "solo", Verdict: "Blocking issues found", Markdown: "x", Findings: []outFinding{
+			{Lens: "security", File: "a.go", Line: 1, Severity: "critical", Title: "rce"},
+		}},
+	}
+	out := renderConsensus(models, clusterFindings(models))
+	headline := out
+	if i := strings.Index(out, "single-model finding"); i >= 0 {
+		headline = out[:i]
+	}
+	if !strings.Contains(headline, "rce") {
+		t.Errorf("lone critical should be in the headline, not folded:\n%s", out)
+	}
+}
+
+func TestClusterSlidingWindowMergesChain(t *testing.T) {
+	// Findings at 10, 13, 16 (each 3 apart) from three models must merge into ONE
+	// cluster — the window slides with the span instead of anchoring at line 10.
+	models := []modelFindings{
+		{Model: "m1", Findings: []outFinding{{Lens: "x", File: "a.go", Line: 10, Severity: "medium", Title: "t"}}},
+		{Model: "m2", Findings: []outFinding{{Lens: "x", File: "a.go", Line: 13, Severity: "medium", Title: "t"}}},
+		{Model: "m3", Findings: []outFinding{{Lens: "x", File: "a.go", Line: 16, Severity: "medium", Title: "t"}}},
+	}
+	clusters := clusterFindings(models)
+	if len(clusters) != 1 {
+		t.Fatalf("chain 10/13/16 should merge into 1 cluster, got %d", len(clusters))
+	}
+	if len(clusters[0].models) != 3 {
+		t.Errorf("want 3 models in the merged cluster, got %d", len(clusters[0].models))
+	}
+}
+
+func TestRenderConsensusExcludesErroredFromDenominator(t *testing.T) {
+	models := []modelFindings{
+		{Model: "m1", Verdict: "Minor issues", Markdown: "a", Findings: []outFinding{
+			{Lens: "security", File: "a.go", Line: 9, Severity: "medium", Title: "leak"}}},
+		{Model: "m2", Verdict: "Minor issues", Markdown: "b", Findings: []outFinding{
+			{Lens: "security", File: "a.go", Line: 9, Severity: "medium", Title: "leak"}}},
+		{Model: "broken", Verdict: "reviewer failed", Errored: true, Markdown: "boom"},
+	}
+	out := renderConsensus(models, clusterFindings(models))
+	// Denominator is the 2 effective models, not 3; the failure is noted.
+	if !strings.Contains(out, "2/2") {
+		t.Errorf("errored model must be excluded from the denominator (want 2/2):\n%s", out)
+	}
+	if !strings.Contains(out, "1 failed") {
+		t.Errorf("expected a '1 failed' note:\n%s", out)
+	}
+	if !strings.Contains(out, "reviewer failed") {
+		t.Errorf("errored model should still appear (folded) as failed:\n%s", out)
+	}
+}
+
+func TestRenderConsensusLoneHighFolds(t *testing.T) {
+	// A single-model HIGH (not critical) folds — only consensus or a lone CRITICAL
+	// earns the headline, so a lone Blocking-lens finding doesn't reintroduce noise.
+	models := []modelFindings{
+		{Model: "solo", Verdict: "Blocking issues found", Markdown: "x", Findings: []outFinding{
+			{Lens: "security", File: "a.go", Line: 1, Severity: "high", Title: "maybe-bug"}}},
+	}
+	out := renderConsensus(models, clusterFindings(models))
+	head := out
+	if i := strings.Index(out, "single-model finding"); i >= 0 {
+		head = out[:i]
+	}
+	if strings.Contains(head, "maybe-bug") {
+		t.Errorf("a lone HIGH should fold, not headline:\n%s", out)
+	}
+}
+
+func TestWriteAndConsolidateRoundTrip(t *testing.T) {
+	dir := t.TempDir()
+
+	// Two model artifacts on disk.
+	write := func(name string, mf modelFindings) {
+		data, _ := json.Marshal(mf)
+		if err := os.WriteFile(filepath.Join(dir, name), data, 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	write("m1.json", modelFindings{Model: "m1", Provider: "ollama", Verdict: "Minor issues", Markdown: "md1",
+		Findings: []outFinding{{Lens: "security", File: "x.go", Line: 3, Severity: "medium", Title: "leak"}}})
+	write("m2.json", modelFindings{Model: "m2", Provider: "ollama", Verdict: "Minor issues", Markdown: "md2",
+		Findings: []outFinding{{Lens: "security", File: "x.go", Line: 3, Severity: "high", Title: "leak"}}})
+	// A junk file must be skipped, not crash consolidation.
+	if err := os.WriteFile(filepath.Join(dir, "notes.txt"), []byte("ignore me"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	t.Setenv("GADFLY_CONSOLIDATE_DIR", dir)
+	// runConsolidate prints to stdout; capture it.
+	out := captureStdout(t, func() {
+		if err := runConsolidate(); err != nil {
+			t.Fatalf("runConsolidate: %v", err)
+		}
+	})
+	if !strings.HasPrefix(strings.TrimSpace(out), consensusMarker) {
+		t.Errorf("consolidated output must lead with the marker:\n%s", out)
+	}
+	if !strings.Contains(out, "2/2") || !strings.Contains(out, "x.go:3") {
+		t.Errorf("expected the agreed x.go:3 finding at 2/2:\n%s", out)
+	}
+}
+
+func TestRunConsolidateEmptyDirErrors(t *testing.T) {
+	t.Setenv("GADFLY_CONSOLIDATE_DIR", t.TempDir())
+	if err := runConsolidate(); err == nil {
+		t.Error("want an error for an empty consolidate dir (entrypoint falls back)")
+	}
+}
+
+// captureStdout redirects os.Stdout for the duration of fn and returns what was
+// written.
+func captureStdout(t *testing.T, fn func()) string {
+	t.Helper()
+	orig := os.Stdout
+	r, w, err := os.Pipe()
+	if err != nil {
+		t.Fatal(err)
+	}
+	os.Stdout = w
+	defer func() { os.Stdout = orig }() // restore even if fn panics
+	done := make(chan string)
+	go func() {
+		var sb strings.Builder
+		buf := make([]byte, 4096)
+		for {
+			n, err := r.Read(buf)
+			if n > 0 {
+				sb.Write(buf[:n])
+			}
+			if err != nil {
+				break
+			}
+		}
+		r.Close()
+		done <- sb.String()
+	}()
+	fn()
+	w.Close()
+	return <-done
+}

cmd/gadfly/consolidate.go

@@ -28,6 +28,22 @@ func (v verdict) label() string {
 	}
 }
 
+// severity maps a verdict to a canonical severity word (the same vocabulary as
+// the structured findings: critical/high/medium/small/trivial). Used as the
+// raw_severity for heuristic-scraped findings, which carry no per-finding
+// severity of their own — so the telemetry store sees one consistent vocabulary
+// instead of a mix of canonical words and full verdict phrases.
+func (v verdict) severity() string {
+	switch v {
+	case verdictBlocking:
+		return "high"
+	case verdictMinor:
+		return "small"
+	default:
+		return "trivial"
+	}
+}
+
 // parseVerdict extracts a specialist's self-reported verdict from its output.
 // The base prompt tells each lens to lead with one of the three phrases.
 func parseVerdict(out string) verdict {
@@ -100,7 +116,9 @@ func renderConsolidated(results []specialistResult) string {
 		headline, len(results), strings.Join(specialistNames(results), ", "))
 
 	for _, r := range results {
-		body := strings.TrimSpace(r.out)
+		// Strip the machine-readable ```gadfly-findings block — it's for tooling
+		// (telemetry/consolidation), not for human readers of the comment.
+		body := strings.TrimSpace(stripFindingsBlock(r.out))
 		if body == "" {
 			body = "_(no output)_"
 		}

cmd/gadfly/emit.go

@@ -10,11 +10,14 @@ package main
 //     the review markdown consumed by run.sh) and never changes the exit code.
 //   - It depends only on the Go stdlib (net/http).
 //
-// Findings are extracted heuristically from each lens's markdown: a "path:line"
+// Findings come PRIMARILY from each lens's machine-readable ```gadfly-findings
-// reference (e.g. run/executor.go:166) anchors a finding, whose title is the
+// block (findings.go: exact file/line + per-finding severity/confidence). When a
-// nearest preceding markdown heading / numbered item / bold lead-in (else the
+// model emits no parseable block, emit falls back to a heuristic prose scrape: a
-// first sentence of the finding's paragraph). This is best-effort signal for the
+// "path:line" reference (e.g. run/executor.go:166) anchors a finding, whose title
-// store to aggregate — not a structured contract the reviewer guarantees.
+// is the nearest preceding markdown heading / numbered item / bold lead-in (else
+// the first sentence of the finding's paragraph). The scrape is best-effort
+// signal for the store to aggregate, not a structured contract the reviewer
+// guarantees.
 
 import (
 	"bytes"
@@ -50,11 +53,16 @@ var pathLineRe = regexp.MustCompile(`([A-Za-z0-9_./-]+\.[A-Za-z0-9]+):(\d+)`)
 var numberedRe = regexp.MustCompile(`^\d+[.)]\s+(.+)$`)
 
 // finding is one extracted issue: where it points and a derived human title.
+// severity/confidence are populated from the structured ```gadfly-findings block
+// (extractStructuredFindings); they are empty for findings recovered by the
+// heuristic prose scrape (parseFindings), which has no per-finding signal.
 type finding struct {
 	file       string
 	line       int
 	title      string
 	detail     string
+	severity   string // per-finding: critical/high/medium/small/trivial ("" if heuristic)
+	confidence string // per-finding: high/medium/low ("" if heuristic)
 }
 
 // runPayload is the POST /runs body. Field names match the gadfly-reports API EXACTLY.
@@ -86,6 +94,7 @@ type reportPayload struct {
 	Provider    string `json:"provider"`
 	RunID       string `json:"run_id"`
 	RawSeverity string `json:"raw_severity"`
+	Confidence  string `json:"confidence"` // per-finding high/medium/low ("" if heuristic)
 	Detail      string `json:"detail"`
 }
 
@@ -116,21 +125,12 @@ func emit(results []specialistResult, elapsed time.Duration) {
 		// InputTokens/OutputTokens/CostUSD stay nil -> JSON null (not metered).
 	}
 
+	// collectFindings (consensus.go) skips clean/errored lenses and fills in each
+	// finding's severity (per-finding when structured, else derived from the lens
+	// verdict), so every reportPayload carries a canonical raw_severity.
 	var reports []reportPayload
 	for _, r := range results {
-		if r.errored {
+		for _, f := range collectFindings(r) {
-			continue // a failed lens contributes no findings
-		}
-		// A lens that reports "No material issues found" has nothing to flag —
-		// its path:line references are verification notes ("verified X at
-		// file:line is safe"), not problems. Extracting them pollutes the
-		// findings store with false positives and unfairly penalizes thorough
-		// reviewers that do clean passes, so a clean lens emits no findings.
-		if r.verdict == verdictClean {
-			continue
-		}
-		sev := r.verdict.label()
-		for _, f := range parseFindings(r.spec, r.out) {
 			reports = append(reports, reportPayload{
 				Repo:        repo,
 				PR:          pr,
@@ -141,7 +141,8 @@ func emit(results []specialistResult, elapsed time.Duration) {
 				Model:       model,
 				Provider:    provider,
 				RunID:       runID,
-				RawSeverity: sev,
+				RawSeverity: f.severity,
+				Confidence:  f.confidence,
 				Detail:      f.detail,
 			})
 		}

cmd/gadfly/emit_test.go

@@ -157,7 +157,7 @@ func TestEmit_PostsRunsAndReports(t *testing.T) {
 	if len(reportBody) != 2 {
 		t.Fatalf("/reports array length = %d, want 2", len(reportBody))
 	}
-	for _, k := range []string{"repo", "pr", "lens", "file", "line", "title", "model", "provider", "run_id", "raw_severity", "detail"} {
+	for _, k := range []string{"repo", "pr", "lens", "file", "line", "title", "model", "provider", "run_id", "raw_severity", "confidence", "detail"} {
 		if _, ok := reportBody[0][k]; !ok {
 			t.Errorf("/reports[0] missing field %q (got keys %v)", k, keysOf(reportBody[0]))
 		}
@@ -171,8 +171,11 @@ func TestEmit_PostsRunsAndReports(t *testing.T) {
 	if reportBody[0]["line"] != float64(166) {
 		t.Errorf("reports[0].line = %v, want 166", reportBody[0]["line"])
 	}
-	if reportBody[0]["raw_severity"] != "Blocking issues found" {
+	// No structured block in sampleLensMarkdown, so this is a heuristic-scraped
-		t.Errorf("reports[0].raw_severity = %v, want 'Blocking issues found'", reportBody[0]["raw_severity"])
+	// finding: raw_severity is the canonical word derived from the lens verdict
+	// (Blocking -> "high"), not the full verdict phrase.
+	if reportBody[0]["raw_severity"] != "high" {
+		t.Errorf("reports[0].raw_severity = %v, want 'high'", reportBody[0]["raw_severity"])
 	}
 	if reportBody[0]["run_id"] != "owner/repo#7:ollama-cloud/qwen3" {
 		t.Errorf("reports[0].run_id = %v, want owner/repo#7:ollama-cloud/qwen3", reportBody[0]["run_id"])

cmd/gadfly/engine.go

@@ -33,14 +33,18 @@ type reviewEngine interface {
 	runPass(ctx context.Context, system, task string, maxSteps int) (string, error)
 }
 
-// majordomoEngine drives the in-process majordomo agent over the repo sandbox.
+// majordomoEngine drives the in-process review path over the repo sandbox. It no
+// longer calls majordomo's agent loop directly: each pass runs through an
+// executus run.Executor (see executus.go), which adds context compaction, run
+// bounding, the per-PR budget gate, and the wrap-up critic. mdl is retained only
+// so auto-select can fall back to the review model as its selector.
 type majordomoEngine struct {
+	rex *reviewExecutor
 	mdl llm.Model
-	fsTools *repoFS
 }
 
 func (e *majordomoEngine) runPass(ctx context.Context, system, task string, maxSteps int) (string, error) {
-	return runAgent(ctx, e.mdl, e.fsTools, system, task, maxSteps)
+	return e.rex.run(ctx, system, task, maxSteps)
 }
 
 // claudeCodeEngine reviews by shelling out to the `claude` CLI (Claude Code) in

cmd/gadfly/executus.go

@@ -0,0 +1,371 @@
+package main
+
+// executus.go wires gadfly's agentic review path onto the executus run kernel
+// (gitea.stevedudenhoeffer.com/steve/executus), layered above majordomo. The
+// majordomoEngine no longer drives majordomo's agent loop directly; it builds a
+// run.Executor that gives gadfly, for free:
+//
+//   - context compaction (executus/compact): once the transcript a step would
+//     SEND crosses a token threshold derived from the model's real context
+//     window, the runaway middle is folded into a one-paragraph summary by a
+//     cheap summarizer model — so a big diff + accumulating read_file/grep
+//     results can't balloon every re-sent step (the large-PR burn).
+//   - run bounding + a per-PR spend budget (executus/run Ports.Budget): a hard
+//     token/seconds ceiling so a pathological PR can't drain the usage block.
+//   - the wrap-up nudge, re-expressed as an executus Critic (Ports.Critic): the
+//     steer that tells a step-hungry model to stop investigating and write its
+//     answer is now the critic seam, not a bespoke RunOption.
+//
+// Everything degrades to today's behavior when unconfigured: nil summarizer or a
+// 0 context window disables compaction; nil budget disables the ceiling; the
+// claude-code engine shells out and is unaffected by any of this.
+//
+// gadfly keeps its own model.go resolution (so GADFLY_ENDPOINT_<NAME> http
+// aliases, failover chains, and the claude-code engine all survive) — the
+// run.Executor is handed gadfly's already-resolved model via a trivial resolver,
+// not routed through executus's tier table.
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"strconv"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	llm "gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+
+	"gitea.stevedudenhoeffer.com/steve/executus/compact"
+	"gitea.stevedudenhoeffer.com/steve/executus/model"
+	exrun "gitea.stevedudenhoeffer.com/steve/executus/run"
+	exectool "gitea.stevedudenhoeffer.com/steve/executus/tool"
+)
+
+const (
+	// defaultCompactRatio is the fraction of the model's context window at which
+	// compaction fires. It is deliberately LOWER than executus's own 0.7 default:
+	// on the large-PR burn the per-step transcript is the embedded diff (~17K) plus
+	// accumulating read_file results, which rarely reaches 0.7×262K≈183K — so a
+	// 0.7 threshold never bites. ~0.45×262K≈118K folds the runaway middle while a
+	// transcript is still well under the cap. Override with GADFLY_COMPACT_RATIO.
+	defaultCompactRatio = 0.45
+	// defaultCompactKeepRecent / defaultCompactSummaryWords mirror executus's own
+	// compactor defaults; surfaced as gadfly env knobs for tuning.
+	defaultCompactKeepRecent   = 8
+	defaultCompactSummaryWords = 200
+	// contextTokenLookupTimeout bounds the one-shot /api/show call that resolves a
+	// cloud model's context window at executor-build time. Kept short so a slow or
+	// unreachable endpoint adds at most this to startup before degrading to
+	// no-compaction (rather than the provider cache's default 15s).
+	contextTokenLookupTimeout = 5 * time.Second
+)
+
+// runSeq mints a unique-per-process RunID suffix for each executor run so audit
+// and the run kernel can tell one pass from another within a binary process.
+var runSeq atomic.Uint64
+
+// wrappedTool adapts an already-built majordomo llm.Tool (gadfly's sandboxed
+// read_file/grep/get_diff/… closures over the repoFS) to executus's tool.Tool
+// interface so the run kernel can build a toolbox from them by name. gadfly's
+// tools need no caller/channel identity, so BuildLLM ignores the Invocation and
+// returns the pre-built tool; Permission is the zero value (private, ungated).
+type wrappedTool struct{ t llm.Tool }
+
+func (w wrappedTool) Name() string                            { return w.t.Name }
+func (w wrappedTool) Description() string                     { return w.t.Description }
+func (w wrappedTool) Permission() exectool.Permission         { return exectool.Permission{} }
+func (w wrappedTool) BuildLLM(_ exectool.Invocation) llm.Tool { return w.t }
+
+// gadflyToolRegistry registers the repo's read-only tools (plus the optional
+// delegate_investigation worker tool) in a fresh executus tool.Registry and
+// returns it along with the tool names for RunnableAgent.LowLevelTools.
+func gadflyToolRegistry(fs *repoFS) (exectool.Registry, []string, error) {
+	reg := exectool.NewRegistry()
+	tools := fs.allTools()
+	names := make([]string, 0, len(tools))
+	for _, t := range tools {
+		if err := reg.Register(wrappedTool{t: t}); err != nil {
+			return nil, nil, fmt.Errorf("register tool %q: %w", t.Name, err)
+		}
+		names = append(names, t.Name)
+	}
+	return reg, names, nil
+}
+
+// gadflyBudget is gadfly's per-PR spend ceiling, satisfying run.Ports.Budget.
+// It gates a run BEFORE it makes any model call (Check) once the process has
+// spent its token or wall-clock allowance on this PR. Tokens are fed in
+// out-of-band via addUsage (the Budget interface's Commit only carries seconds);
+// the engine calls addUsage after each pass with run.Result.Usage. A nil
+// *gadflyBudget is never installed — caps of 0 mean "unlimited", so the port is
+// only wired when at least one cap is set.
+//
+// The guard is PASS-granular: Check runs before each pass, so it stops the NEXT
+// pass once the budget is spent but cannot abort a single runaway pass mid-flight.
+// The swarm-wide GADFLY_PR_BUDGET_SECS wall-clock backstop (entrypoint.sh) is what
+// bounds a mid-pass runaway.
+type gadflyBudget struct {
+	mu         sync.Mutex
+	maxTokens  int64
+	maxSeconds float64
+	tokens     int64
+	seconds    float64
+}
+
+// newPRBudget builds the per-PR budget from env, or nil when neither cap is set
+// (the default — the swarm-wide ceiling lives in entrypoint.sh; this is the
+// per-process belt to its suspenders).
+func newPRBudget() *gadflyBudget {
+	toks := envInt("GADFLY_PR_TOKEN_BUDGET", 0)
+	secs := envInt("GADFLY_PR_TIME_BUDGET_SECS", 0)
+	if toks <= 0 && secs <= 0 {
+		return nil
+	}
+	return &gadflyBudget{maxTokens: int64(toks), maxSeconds: float64(secs)}
+}
+
+func (b *gadflyBudget) Check(_ context.Context, _ string) error {
+	if b == nil {
+		return nil
+	}
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	if b.maxTokens > 0 && b.tokens >= b.maxTokens {
+		return fmt.Errorf("gadfly: per-PR token budget exhausted (%d/%d)", b.tokens, b.maxTokens)
+	}
+	if b.maxSeconds > 0 && b.seconds >= b.maxSeconds {
+		return fmt.Errorf("gadfly: per-PR time budget exhausted (%.0f/%.0fs)", b.seconds, b.maxSeconds)
+	}
+	return nil
+}
+
+func (b *gadflyBudget) Commit(_ context.Context, _ string, runtimeSeconds float64) {
+	if b == nil {
+		return
+	}
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	b.seconds += runtimeSeconds
+}
+
+// addUsage records a finished pass's token spend toward the budget. Safe on nil.
+func (b *gadflyBudget) addUsage(u llm.Usage) {
+	if b == nil {
+		return
+	}
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	b.tokens += int64(u.InputTokens) + int64(u.OutputTokens)
+}
+
+// wrapUpCritic re-expresses gadfly's wrap-up nudge as an executus run.Critic:
+// once a run comes within wrapUpReserve steps of its cap, Steer() injects the
+// "stop calling tools and write your final answer" message so a thorough model
+// spends its last steps finalizing instead of hard-failing empty. It sets no
+// hard deadline (Deadline()==zero) and never raises the step ceiling
+// (MaxSteps()==0, defer to the run's MaxIterations) — it is purely the nudge.
+type wrapUpCritic struct{ reserve int }
+
+func (c *wrapUpCritic) Monitor(_ context.Context, info exrun.RunInfo, _ time.Duration) exrun.CriticHandle {
+	return &wrapUpHandle{maxSteps: info.MaxIterations, reserve: c.reserve}
+}
+
+type wrapUpHandle struct {
+	mu       sync.Mutex
+	maxSteps int
+	reserve  int
+	done     int // steps completed so far
+	nudged   bool
+}
+
+func (h *wrapUpHandle) RecordStep(iter int, _ *llm.Response) {
+	h.mu.Lock()
+	h.done = iter + 1
+	h.mu.Unlock()
+}
+func (h *wrapUpHandle) RecordToolStart(string, string) {}
+func (h *wrapUpHandle) Steer() []llm.Message {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	at := h.maxSteps - h.reserve
+	if at < 1 {
+		at = 1
+	}
+	if !h.nudged && h.maxSteps > 0 && h.done >= at {
+		h.nudged = true
+		return []llm.Message{llm.UserText(wrapUpInstruction)}
+	}
+	return nil
+}
+func (h *wrapUpHandle) Deadline() time.Time { return time.Time{} }
+func (h *wrapUpHandle) MaxSteps() int       { return 0 }
+func (h *wrapUpHandle) KillCause() error    { return nil }
+func (h *wrapUpHandle) Stop()               {}
+
+// reviewExecutor bundles a run.Executor with the per-run wiring the engine needs
+// for each pass (the tool names to expose, the model spec to report as the tier,
+// the per-PR caller id, and the budget to feed token usage into).
+type reviewExecutor struct {
+	ex        *exrun.Executor
+	toolNames []string
+	modelSpec string
+	callerID  string
+	budget    *gadflyBudget
+}
+
+// newReviewExecutor builds the run.Executor for the in-process majordomo review
+// path. mdl is gadfly's already-resolved review model; summarizer is the cheap
+// model the compactor uses (nil disables compaction). Compaction also needs the
+// model's context window (resolved once here, not per pass); a 0 window likewise
+// disables it. The budget (may be nil) becomes the run.Ports.Budget gate.
+func newReviewExecutor(fs *repoFS, mdl, summarizer llm.Model, modelSpec string, budget *gadflyBudget) (*reviewExecutor, error) {
+	reg, names, err := gadflyToolRegistry(fs)
+	if err != nil {
+		return nil, err
+	}
+
+	// gadfly resolves exactly one model per process; the run kernel's resolver
+	// just hands that model back regardless of the tier string it is asked for.
+	modelsResolver := func(ctx context.Context, _ string) (context.Context, llm.Model, error) {
+		return ctx, mdl, nil
+	}
+
+	var compactor compact.CompactorFactory
+	var ctxTokens func(string) int
+	if summarizer != nil && compactionEnabled() {
+		if window := resolveContextTokens(modelSpec); window > 0 {
+			sumResolver := func(ctx context.Context, _ string) (context.Context, llm.Model, error) {
+				return ctx, summarizer, nil
+			}
+			compactor = compact.NewCompactor(compact.CompactorConfig{
+				Models:         sumResolver,
+				KeepRecent:     envInt("GADFLY_COMPACT_KEEP_RECENT", defaultCompactKeepRecent),
+				SummaryWordCap: envInt("GADFLY_COMPACT_SUMMARY_WORDS", defaultCompactSummaryWords),
+			})
+			ctxTokens = func(string) int { return window } // memoized: one window per process
+		}
+	}
+
+	var ports exrun.Ports
+	ports.Critic = &wrapUpCritic{reserve: wrapUpReserve()}
+	if budget != nil {
+		ports.Budget = budget
+	}
+
+	cfg := exrun.Config{
+		Registry:      reg,
+		Models:        modelsResolver,
+		Compactor:     compactor,
+		ContextTokens: ctxTokens,
+		Defaults: exrun.Defaults{
+			// MaxIterations/MaxRuntime are intentionally omitted: every pass sets its
+			// own per-run cap on the RunnableAgent below (the review and recheck caps
+			// differ), so a Defaults value here would always be overridden — dead. This
+			// leaves only the cross-pass guards + the compaction ratio.
+			MaxConsecutiveToolErrors: 4,
+			MaxSameToolCallRepeats:   4,
+			CompactionThresholdRatio: compactRatio(),
+			FallbackTier:             modelSpec,
+		},
+		Ports: ports,
+	}
+	return &reviewExecutor{
+		ex:        exrun.New(cfg),
+		toolNames: names,
+		modelSpec: modelSpec,
+		callerID:  prCallerID(),
+		budget:    budget,
+	}, nil
+}
+
+// run executes one agent pass (review or recheck) through the run kernel and
+// returns the model's final text. An empty answer with no error is reported as
+// an error so the caller (reviewWithSpecialist) renders the advisory "reviewer
+// failed to complete" notice rather than a blank section.
+func (r *reviewExecutor) run(ctx context.Context, system, task string, maxSteps int) (string, error) {
+	res := r.ex.Run(ctx, exrun.RunnableAgent{
+		Name:          "gadfly-review",
+		SystemPrompt:  system,
+		ModelTier:     r.modelSpec,
+		MaxIterations: maxSteps,
+		MaxRuntime:    reviewTimeout(),
+		LowLevelTools: r.toolNames,
+		Critic:        exrun.CriticConfig{Enabled: true},
+	}, exectool.Invocation{
+		RunID:    fmt.Sprintf("gadfly-%d", runSeq.Add(1)),
+		CallerID: r.callerID,
+	}, task)
+
+	// Feed token spend toward the per-PR budget out-of-band (Commit carries only
+	// seconds; the executor already called it). Safe on a nil budget.
+	r.budget.addUsage(res.Usage)
+
+	if res.Err != nil {
+		return "", res.Err
+	}
+	if out := strings.TrimSpace(res.Output); out != "" {
+		return out, nil
+	}
+	return "", errors.New("agent produced no output")
+}
+
+// prCallerID is the budget/audit caller key: the repo + PR, so a budget keyed on
+// it is naturally per-PR. Falls back to "local" for an out-of-CI run.
+func prCallerID() string {
+	repo := strings.TrimSpace(os.Getenv("GADFLY_REPO"))
+	pr := strings.TrimSpace(os.Getenv("GADFLY_PR"))
+	if repo == "" && pr == "" {
+		return "local"
+	}
+	return repo + "#" + pr
+}
+
+// compactionEnabled reports whether context compaction should be wired. On
+// unless GADFLY_COMPACT is explicitly falsey.
+func compactionEnabled() bool { return envBool("GADFLY_COMPACT", true) }
+
+// compactRatio is the compaction threshold as a fraction of the model context
+// window (GADFLY_COMPACT_RATIO), clamped to (0,1]; default defaultCompactRatio.
+func compactRatio() float64 {
+	v := strings.TrimSpace(os.Getenv("GADFLY_COMPACT_RATIO"))
+	if v == "" {
+		return defaultCompactRatio
+	}
+	f, err := strconv.ParseFloat(v, 64)
+	if err != nil || f <= 0 || f > 1 {
+		return defaultCompactRatio
+	}
+	return f
+}
+
+// resolveContextTokens returns the review model's context window in tokens, used
+// to set the compaction threshold. GADFLY_MODEL_CONTEXT_TOKENS overrides it
+// (needed for custom/self-hosted endpoints executus can't introspect); otherwise
+// it asks executus/model, which knows the static catalog and can fetch an
+// Ollama Cloud model's limit via /api/show (one call, at executor-build time).
+// Returns 0 — disabling compaction — for an unknown model, mirroring executus's
+// "unknown ⇒ don't budget" contract.
+func resolveContextTokens(modelSpec string) int {
+	if v := envInt("GADFLY_MODEL_CONTEXT_TOKENS", 0); v > 0 {
+		return v
+	}
+	key := strings.TrimSpace(os.Getenv("GADFLY_API_KEY"))
+	if key == "" {
+		key = strings.TrimSpace(os.Getenv("OLLAMA_API_KEY"))
+	}
+	cache := model.NewCloudOllamaLimitCache("", key, nil)
+	ctx, cancel := context.WithTimeout(context.Background(), contextTokenLookupTimeout)
+	defer cancel()
+	if n, ok := model.MaxContextTokensResolving(ctx, modelSpec, cache); ok {
+		return n
+	}
+	// Unknown model or a failed lookup (e.g. no key / unreachable endpoint): don't
+	// guess — compaction is disabled. Log it so a misconfiguration is debuggable
+	// rather than silently dropping the protection. Set GADFLY_MODEL_CONTEXT_TOKENS
+	// to force a window for an endpoint executus can't introspect.
+	fmt.Fprintf(os.Stderr, "gadfly: no context window resolved for %q; compaction disabled (set GADFLY_MODEL_CONTEXT_TOKENS to enable it)\n", modelSpec)
+	return 0
+}

cmd/gadfly/executus_test.go

@@ -0,0 +1,140 @@
+package main
+
+import (
+	"context"
+	"testing"
+
+	llm "gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+func TestGadflyBudget(t *testing.T) {
+	ctx := context.Background()
+
+	// A nil budget never blocks and never panics.
+	var nilB *gadflyBudget
+	if err := nilB.Check(ctx, "pr"); err != nil {
+		t.Errorf("nil budget Check should be nil, got %v", err)
+	}
+	nilB.Commit(ctx, "pr", 10)
+	nilB.addUsage(llm.Usage{InputTokens: 5})
+
+	// Token ceiling: passes until usage crosses it.
+	b := &gadflyBudget{maxTokens: 100}
+	if err := b.Check(ctx, "pr"); err != nil {
+		t.Fatalf("fresh budget should pass, got %v", err)
+	}
+	b.addUsage(llm.Usage{InputTokens: 60, OutputTokens: 50}) // 110 >= 100
+	if err := b.Check(ctx, "pr"); err == nil {
+		t.Error("budget over the token cap should reject the next run")
+	}
+
+	// Seconds ceiling, accumulated via Commit.
+	s := &gadflyBudget{maxSeconds: 30}
+	s.Commit(ctx, "pr", 31)
+	if err := s.Check(ctx, "pr"); err == nil {
+		t.Error("budget over the time cap should reject the next run")
+	}
+}
+
+func TestNewPRBudget(t *testing.T) {
+	t.Setenv("GADFLY_PR_TOKEN_BUDGET", "")
+	t.Setenv("GADFLY_PR_TIME_BUDGET_SECS", "")
+	if newPRBudget() != nil {
+		t.Error("no caps set should yield a nil (disabled) budget")
+	}
+	t.Setenv("GADFLY_PR_TOKEN_BUDGET", "1000")
+	if b := newPRBudget(); b == nil || b.maxTokens != 1000 {
+		t.Errorf("token cap should build a budget with maxTokens=1000, got %+v", b)
+	}
+}
+
+func TestCompactRatio(t *testing.T) {
+	t.Setenv("GADFLY_COMPACT_RATIO", "")
+	if got := compactRatio(); got != defaultCompactRatio {
+		t.Errorf("default ratio = %v, want %v", got, defaultCompactRatio)
+	}
+	t.Setenv("GADFLY_COMPACT_RATIO", "0.6")
+	if got := compactRatio(); got != 0.6 {
+		t.Errorf("ratio override = %v, want 0.6", got)
+	}
+	for _, bad := range []string{"0", "-1", "2", "nope"} {
+		t.Setenv("GADFLY_COMPACT_RATIO", bad)
+		if got := compactRatio(); got != defaultCompactRatio {
+			t.Errorf("invalid ratio %q should fall back to the default, got %v", bad, got)
+		}
+	}
+}
+
+func TestResolveContextTokensOverride(t *testing.T) {
+	// The explicit override short-circuits any model introspection (no network).
+	t.Setenv("GADFLY_MODEL_CONTEXT_TOKENS", "123456")
+	if got := resolveContextTokens("anything"); got != 123456 {
+		t.Errorf("explicit context-token override = %d, want 123456", got)
+	}
+}
+
+func TestPRCallerID(t *testing.T) {
+	t.Setenv("GADFLY_REPO", "")
+	t.Setenv("GADFLY_PR", "")
+	if got := prCallerID(); got != "local" {
+		t.Errorf("no repo/PR should be %q, got %q", "local", got)
+	}
+	t.Setenv("GADFLY_REPO", "steve/mort")
+	t.Setenv("GADFLY_PR", "1367")
+	if got := prCallerID(); got != "steve/mort#1367" {
+		t.Errorf("callerID = %q, want steve/mort#1367", got)
+	}
+}
+
+func TestCompactionEnabled(t *testing.T) {
+	for _, v := range []string{"", "1", "true", "yes"} {
+		t.Setenv("GADFLY_COMPACT", v)
+		if !compactionEnabled() {
+			t.Errorf("GADFLY_COMPACT=%q should be enabled", v)
+		}
+	}
+	for _, v := range []string{"0", "false", "no", "off"} {
+		t.Setenv("GADFLY_COMPACT", v)
+		if compactionEnabled() {
+			t.Errorf("GADFLY_COMPACT=%q should be disabled", v)
+		}
+	}
+}
+
+func TestGadflyToolRegistry(t *testing.T) {
+	fs, err := newRepoFS(t.TempDir(), "diff")
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, names, err := gadflyToolRegistry(fs)
+	if err != nil {
+		t.Fatalf("gadflyToolRegistry: %v", err)
+	}
+	want := map[string]bool{"read_file": true, "list_dir": true, "grep": true, "find_files": true, "get_diff": true}
+	has := func(n string) bool {
+		for _, g := range names {
+			if g == n {
+				return true
+			}
+		}
+		return false
+	}
+	for n := range want {
+		if !has(n) {
+			t.Errorf("registry missing tool %q (got %v)", n, names)
+		}
+	}
+	if has("delegate_investigation") {
+		t.Error("delegate_investigation must be absent without a worker model")
+	}
+
+	// With a worker model the delegate tool is registered too.
+	fs.worker = fakeModel(t, "x")
+	_, names, err = gadflyToolRegistry(fs)
+	if err != nil {
+		t.Fatalf("gadflyToolRegistry with worker: %v", err)
+	}
+	if !has("delegate_investigation") {
+		t.Errorf("delegate_investigation should be registered with a worker model, got %v", names)
+	}
+}

cmd/gadfly/findings.go

@@ -0,0 +1,236 @@
+package main
+
+// Structured findings: the machine-readable contract between a lens's output and
+// the telemetry/consolidation pipeline. Each lens is asked (see
+// scripts/system-prompt.txt) to append a fenced ```gadfly-findings code block
+// holding a JSON array of its findings. Parsing that exact block is far more
+// reliable than scraping prose with a path:line regex (emit.go's heuristic),
+// and it carries PER-FINDING severity + confidence the prose verdict can't.
+//
+// Everything here degrades gracefully: a missing, unterminated, or malformed
+// block makes extractStructuredFindings return ok=false (and yield no findings),
+// so the caller falls back to the heuristic scrape — a weak model that ignores
+// the contract still contributes findings, exactly as before.
+
+import (
+	"encoding/json"
+	"strconv"
+	"strings"
+)
+
+// structuredFinding mirrors one element of the ```gadfly-findings JSON array.
+// Line is json.Number so we tolerate both 123 and "123" from less-precise models.
+type structuredFinding struct {
+	File       string      `json:"file"`
+	Line       json.Number `json:"line"`
+	Severity   string      `json:"severity"`
+	Confidence string      `json:"confidence"`
+	Title      string      `json:"title"`
+	Detail     string      `json:"detail"` // optional; the prose paragraph is used when absent
+}
+
+// findingsFence is the info-string that tags the machine-readable block.
+const findingsFence = "gadfly-findings"
+
+// extractStructuredFindings parses the ```gadfly-findings JSON block out of a
+// lens's markdown. It returns the findings and ok=true when a TERMINATED block is
+// present AND parses as a JSON array (an empty array is valid "nothing found" —
+// ok is still true). A missing, unterminated, or unparseable block returns
+// ok=false so the caller falls back to the heuristic scrape.
+//
+// Findings are deduped by file:line (keeping the first, matching parseFindings),
+// findings with no usable file are dropped, and each title/detail is backfilled
+// from the prose when the JSON omits it (best of both: exact location + the human
+// context the model already wrote).
+func extractStructuredFindings(out string) ([]finding, bool) {
+	lines := strings.Split(out, "\n")
+	start, end, ok := findingsSpan(lines)
+	if !ok {
+		return nil, false
+	}
+	var raw []structuredFinding
+	if err := json.Unmarshal([]byte(strings.Join(lines[start+1:end], "\n")), &raw); err != nil {
+		return nil, false
+	}
+
+	findings := make([]finding, 0, len(raw))
+	seen := map[string]bool{}
+	var prose map[string]string // built lazily — only if a finding lacks its own detail
+	for _, sf := range raw {
+		file := strings.TrimSpace(sf.File)
+		if file == "" {
+			continue
+		}
+		ln := 0
+		if n, err := strconv.Atoi(strings.TrimSpace(sf.Line.String())); err == nil && n > 0 {
+			ln = n
+		}
+		key := file + ":" + strconv.Itoa(ln)
+		if ln > 0 { // only dedupe concrete locations; unknown-line findings are kept
+			if seen[key] {
+				continue
+			}
+			seen[key] = true
+		}
+		detail := strings.TrimSpace(sf.Detail)
+		if detail == "" && ln > 0 {
+			if prose == nil {
+				prose = proseParagraphs(out)
+			}
+			detail = prose[key]
+		}
+		title := strings.TrimSpace(sf.Title)
+		if title == "" { // never empty: fall back to the prose detail, then the location
+			if detail != "" {
+				title = truncate(detail, 120)
+			} else if ln > 0 {
+				title = key
+			} else {
+				title = file
+			}
+		}
+		findings = append(findings, finding{
+			file:       file,
+			line:       ln,
+			title:      title,
+			detail:     truncate(detail, 500),
+			severity:   normalizeSeverity(sf.Severity),
+			confidence: normalizeConfidence(sf.Confidence),
+		})
+		if len(findings) >= maxFindingsPerLens {
+			break
+		}
+	}
+	return findings, true
+}
+
+// extractStructuredFindingsOrScrape returns a lens's findings, preferring the
+// structured ```gadfly-findings block and falling back to the heuristic prose
+// scrape when the block is absent, unterminated/malformed, OR parsed to zero
+// usable findings (e.g. an empty [] emitted alongside real prose findings).
+// Factored out of emit() so the fallback rule is unit-testable.
+func extractStructuredFindingsOrScrape(r specialistResult) []finding {
+	if fs, _ := extractStructuredFindings(r.out); len(fs) > 0 {
+		return fs
+	}
+	return parseFindings(r.spec, r.out)
+}
+
+// stripFindingsBlock removes every TERMINATED ```gadfly-findings block from out so
+// the machine-readable JSON never shows in the rendered comment. An UNTERMINATED
+// fence is left in place — treating it as a block would swallow the rest of the
+// comment (e.g. when a model's output was truncated mid-block). Trailing
+// whitespace is trimmed.
+func stripFindingsBlock(out string) string {
+	lines := strings.Split(out, "\n")
+	for {
+		start, end, ok := findingsSpan(lines)
+		if !ok {
+			break
+		}
+		lines = append(lines[:start:start], lines[end+1:]...)
+	}
+	return strings.TrimRight(strings.Join(lines, "\n"), "\n")
+}
+
+// findingsSpan returns the [start,end] inclusive line indices of the first
+// TERMINATED ```gadfly-findings block in lines (start = the opening fence, end =
+// the closing fence), or ok=false when there is none or it is unterminated.
+// extract and strip share it so they always agree on what is (and isn't) a block.
+func findingsSpan(lines []string) (start, end int, ok bool) {
+	start = -1
+	for i, ln := range lines {
+		if isFindingsOpen(ln) {
+			start = i
+			break
+		}
+	}
+	if start < 0 {
+		return 0, 0, false
+	}
+	for j := start + 1; j < len(lines); j++ {
+		if isFenceClose(lines[j]) {
+			return start, j, true
+		}
+	}
+	return 0, 0, false // unterminated
+}
+
+// fenceInfo returns the info-string (text after the backticks) of a code-fence
+// line and whether the line opens/closes a fence at all. A bare ``` yields ("",
+// true); ```gadfly-findings yields ("gadfly-findings", true).
+func fenceInfo(line string) (string, bool) {
+	t := strings.TrimSpace(line)
+	if !strings.HasPrefix(t, "```") {
+		return "", false
+	}
+	return strings.TrimSpace(strings.TrimLeft(t, "`")), true
+}
+
+// isFindingsOpen reports whether line opens a ```gadfly-findings block, matching
+// the info-string EXACTLY (not as a substring) so a fence like ```not-findings
+// can't masquerade as ours.
+func isFindingsOpen(line string) bool {
+	info, ok := fenceInfo(line)
+	return ok && info == findingsFence
+}
+
+// isFenceClose reports whether line is a bare closing fence (``` with no info).
+func isFenceClose(line string) bool {
+	info, ok := fenceInfo(line)
+	return ok && info == ""
+}
+
+// proseParagraphs maps "file:line" -> the prose paragraph that first references
+// it, so a structured finding without its own detail can borrow the human
+// context the model already wrote. Built from the markdown OUTSIDE the findings
+// block (the block is JSON, not prose).
+func proseParagraphs(out string) map[string]string {
+	prose := stripFindingsBlock(out)
+	lines := strings.Split(prose, "\n")
+	m := map[string]string{}
+	for _, loc := range pathLineRe.FindAllStringSubmatchIndex(prose, -1) {
+		key := prose[loc[2]:loc[3]] + ":" + prose[loc[4]:loc[5]]
+		if _, dup := m[key]; dup {
+			continue
+		}
+		li := strings.Count(prose[:loc[0]], "\n")
+		m[key] = paragraphAt(lines, li)
+	}
+	return m
+}
+
+// normalizeSeverity maps a model's severity word onto the canonical set
+// (critical/high/medium/small/trivial), accepting common synonyms. An
+// unrecognized value is returned lowercased so the store still sees the raw word.
+func normalizeSeverity(s string) string {
+	switch t := strings.ToLower(strings.TrimSpace(s)); t {
+	case "critical", "crit", "blocker", "blocking":
+		return "critical"
+	case "high", "major", "severe":
+		return "high"
+	case "medium", "moderate":
+		return "medium"
+	case "small", "low", "minor":
+		return "small"
+	case "trivial", "nit", "nitpick", "info", "informational", "style", "cosmetic":
+		return "trivial"
+	default:
+		return t
+	}
+}
+
+// normalizeConfidence maps a model's confidence word onto high/medium/low,
+// accepting common synonyms; an unrecognized value is returned lowercased.
+func normalizeConfidence(s string) string {
+	switch t := strings.ToLower(strings.TrimSpace(s)); t {
+	case "high", "certain", "confirmed", "verified":
+		return "high"
+	case "medium", "med", "moderate":
+		return "medium"
+	case "low", "unsure", "tentative", "unverified", "speculative":
+		return "low"
+	default:
+		return t
+	}
+}

cmd/gadfly/findings_test.go

@@ -0,0 +1,146 @@
+package main
+
+import (
+	"strings"
+	"testing"
+)
+
+const sampleReview = "**Blocking issues found**\n\n" +
+	"- **Unauthenticated endpoint** — `model.go:184` leaks PR content to a third party.\n" +
+	"- A nit about naming in `util.go:12`.\n\n" +
+	"```gadfly-findings\n" +
+	"[\n" +
+	"  {\"file\": \"model.go\", \"line\": 184, \"severity\": \"high\", \"confidence\": \"high\", \"title\": \"Unauthenticated endpoint\"},\n" +
+	"  {\"file\": \"util.go\", \"line\": 12, \"severity\": \"nit\", \"confidence\": \"medium\", \"title\": \"naming\"}\n" +
+	"]\n" +
+	"```\n"
+
+func TestExtractStructuredFindings(t *testing.T) {
+	fs, ok := extractStructuredFindings(sampleReview)
+	if !ok {
+		t.Fatal("expected ok=true for a well-formed block")
+	}
+	if len(fs) != 2 {
+		t.Fatalf("want 2 findings, got %d", len(fs))
+	}
+	if fs[0].file != "model.go" || fs[0].line != 184 || fs[0].severity != "high" || fs[0].confidence != "high" {
+		t.Errorf("finding[0] mismatch: %+v", fs[0])
+	}
+	// "nit" must normalize to the canonical "trivial".
+	if fs[1].severity != "trivial" {
+		t.Errorf("want severity normalized to trivial, got %q", fs[1].severity)
+	}
+	// Detail is borrowed from the prose paragraph referencing the same file:line.
+	if fs[0].detail == "" {
+		t.Error("expected detail borrowed from prose, got empty")
+	}
+}
+
+func TestExtractStructuredFindingsEmptyArray(t *testing.T) {
+	out := "No material issues found\n\n```gadfly-findings\n[]\n```\n"
+	fs, ok := extractStructuredFindings(out)
+	if !ok {
+		t.Fatal("an empty array is a valid block; want ok=true")
+	}
+	if len(fs) != 0 {
+		t.Fatalf("want 0 findings, got %d", len(fs))
+	}
+}
+
+func TestExtractStructuredFindingsFallback(t *testing.T) {
+	// No block at all -> ok=false so the caller uses the heuristic scrape.
+	if _, ok := extractStructuredFindings("Minor issues\n\n- something at `x.go:1`\n"); ok {
+		t.Error("want ok=false when there is no block")
+	}
+	// Malformed JSON -> ok=false (graceful fallback).
+	bad := "Minor issues\n\n```gadfly-findings\n{not json}\n```\n"
+	if _, ok := extractStructuredFindings(bad); ok {
+		t.Error("want ok=false for malformed JSON")
+	}
+}
+
+func TestExtractStructuredFindingsStringLine(t *testing.T) {
+	// Tolerate a quoted line number from a less-precise model.
+	out := "Minor issues\n\n```gadfly-findings\n[{\"file\":\"a.go\",\"line\":\"42\",\"severity\":\"medium\",\"title\":\"x\"}]\n```\n"
+	fs, ok := extractStructuredFindings(out)
+	if !ok || len(fs) != 1 || fs[0].line != 42 {
+		t.Fatalf("want one finding at line 42, got ok=%v %+v", ok, fs)
+	}
+}
+
+func TestStripFindingsBlock(t *testing.T) {
+	stripped := stripFindingsBlock(sampleReview)
+	if strings.Contains(stripped, findingsFence) {
+		t.Errorf("block not stripped: %q", stripped)
+	}
+	// The prose findings must survive.
+	if !strings.Contains(stripped, "Unauthenticated endpoint") || !strings.Contains(stripped, "util.go:12") {
+		t.Errorf("prose lost during strip: %q", stripped)
+	}
+}
+
+func TestStripFindingsBlockUnterminated(t *testing.T) {
+	// A truncated, unterminated block must NOT swallow the prose before it.
+	out := "Minor issues\n\n- real finding at `x.go:1`\n\n```gadfly-findings\n[{\"file\":\"x.go\""
+	got := stripFindingsBlock(out)
+	if !strings.Contains(got, "real finding at `x.go:1`") {
+		t.Errorf("unterminated block swallowed the prose: %q", got)
+	}
+}
+
+func TestStripFindingsBlockNoBlock(t *testing.T) {
+	in := "Minor issues\n\n- finding at `x.go:9`"
+	if out := stripFindingsBlock(in); out != in {
+		t.Errorf("strip changed block-free text: %q != %q", out, in)
+	}
+}
+
+func TestNormalizeSeverity(t *testing.T) {
+	cases := map[string]string{
+		"Critical": "critical", "blocker": "critical",
+		"major": "high", "HIGH": "high",
+		"moderate": "medium",
+		"minor":    "small", "low": "small", // "minor" and "low" both map to small (consistently)
+		"nit": "trivial", "Style": "trivial",
+		"weird": "weird", // unknown passes through, lowercased
+	}
+	for in, want := range cases {
+		if got := normalizeSeverity(in); got != want {
+			t.Errorf("normalizeSeverity(%q) = %q, want %q", in, got, want)
+		}
+	}
+}
+
+func TestNormalizeConfidence(t *testing.T) {
+	cases := map[string]string{
+		"High": "high", "confirmed": "high",
+		"MEDIUM": "medium", "moderate": "medium",
+		"low": "low", "unverified": "low",
+		"hunch": "hunch", // unknown passes through, lowercased
+	}
+	for in, want := range cases {
+		if got := normalizeConfidence(in); got != want {
+			t.Errorf("normalizeConfidence(%q) = %q, want %q", in, got, want)
+		}
+	}
+}
+
+func TestExtractStructuredFindingsFallbackOnEmpty(t *testing.T) {
+	// A non-clean lens that emitted an empty [] but listed prose findings must
+	// fall through to the heuristic scrape, not silently drop everything.
+	out := "Minor issues\n\n- bug at `pkg/a.go:7`\n\n```gadfly-findings\n[]\n```\n"
+	r := specialistResult{spec: Specialist{Name: "correctness"}, out: out, verdict: verdictMinor}
+	fs := extractStructuredFindingsOrScrape(r)
+	if len(fs) == 0 {
+		t.Fatal("empty [] must fall back to the heuristic scrape, got no findings")
+	}
+	if fs[0].file != "pkg/a.go" || fs[0].line != 7 {
+		t.Errorf("heuristic fallback wrong: %+v", fs[0])
+	}
+}
+
+func TestVerdictSeverity(t *testing.T) {
+	if verdictBlocking.severity() != "high" || verdictMinor.severity() != "small" || verdictUnknown.severity() != "trivial" {
+		t.Error("verdict.severity mapping changed unexpectedly")
+	}
+}

cmd/gadfly/lens_concurrency_test.go

@@ -104,7 +104,7 @@ func TestRunSpecialists_FansOut(t *testing.T) {
 	}
 	specs := threeLenses()
 
-	results := runSpecialists(&majordomoEngine{mdl: mdl, fsTools: fs}, "sys", specs, "task", "diff")
+	results := runSpecialists(testEngine(t, mdl, fs), "sys", specs, "task", "diff")
 
 	if got := peak(); got != 3 {
 		t.Errorf("peak concurrent lenses = %d, want 3", got)
@@ -124,7 +124,7 @@ func TestRunSpecialists_SequentialByDefault(t *testing.T) {
 	}
 	specs := threeLenses()
 
-	results := runSpecialists(&majordomoEngine{mdl: mdl, fsTools: fs}, "sys", specs, "task", "diff")
+	results := runSpecialists(testEngine(t, mdl, fs), "sys", specs, "task", "diff")
 
 	if got := peak(); got != 1 {
 		t.Errorf("peak concurrent lenses = %d, want 1 (sequential by default)", got)
@@ -146,7 +146,7 @@ func TestRunSpecialists_PerProviderFanOut(t *testing.T) {
 	}
 	specs := threeLenses()
 
-	results := runSpecialists(&majordomoEngine{mdl: mdl, fsTools: fs}, "sys", specs, "task", "diff")
+	results := runSpecialists(testEngine(t, mdl, fs), "sys", specs, "task", "diff")
 
 	if got := peak(); got != 3 {
 		t.Errorf("peak concurrent lenses = %d, want 3 (m1 per-provider override)", got)

cmd/gadfly/main.go

@@ -39,10 +39,9 @@
 //	GADFLY_TITLE          PR title (optional).
 //	GADFLY_BODY           PR description (optional).
 //	GADFLY_MAX_STEPS      review-pass step cap (optional, default 24).
-//	GADFLY_WRAPUP_RESERVE steps before the cap at which the agent is told to
+//	GADFLY_WRAPUP_RESERVE steps before the cap at which the wrap-up critic nudges
-//	                        stop investigating and write its answer (optional,
+//	                        the agent to stop investigating and write its answer
-//	                        default 4). Plus a tool-free finalization fallback
+//	                        (optional, default 4).
-//	                        guarantees a step-exhausted pass still emits output.
 //	GADFLY_RECHECK        set to 0/false to skip the recheck pass (optional, default on).
 //	GADFLY_RECHECK_MAX_STEPS recheck-pass step cap (optional, default 16).
 //	GADFLY_TIMEOUT_SECS   overall deadline in seconds, shared by both passes (optional, default 300).
@@ -55,8 +54,8 @@
 //	                        lanes as GADFLY_PROVIDER_CONCURRENCY (e.g.
 //	                        "ollama-cloud=3,m1=1"). Wins over GADFLY_LENS_CONCURRENCY
 //	                        for the model's provider; falls back to it otherwise.
-//	GADFLY_MAX_DIFF_CHARS diff chars embedded in the prompt (optional, default 60000;
+//	GADFLY_MAX_DIFF_CHARS diff chars embedded in the review prompt (optional, default 60000;
-//	                        the full diff is always available via the get_diff tool).
+//	                        the full diff is reachable via the paginated get_diff tool).
 //
 // On success it prints the review to stdout and exits 0. On a usage/config or
 // model error it prints a diagnostic to stderr and exits non-zero; run.sh then
@@ -70,11 +69,9 @@ import (
 	"os"
 	"strconv"
 	"strings"
-	"sync"
 	"time"
 
-	"gitea.stevedudenhoeffer.com/steve/majordomo/agent"
+	"gitea.stevedudenhoeffer.com/steve/executus/fanout"
-	llm "gitea.stevedudenhoeffer.com/steve/majordomo/llm"
 )
 
 const (
@@ -107,13 +104,6 @@ const wrapUpInstruction = "⚠️ You are almost out of your investigation budge
 	"Do not begin any new investigation. If a finding could not be confirmed, drop it or mark it explicitly as unverified. " +
 	"Output the review in the required format right now."
 
-// finalizeInstruction is the user message sent on the tool-free fallback pass
-// when the agent exhausted its budget (or tripped a loop guard) without ever
-// producing a final answer. It forces the model to synthesize whatever it has.
-const finalizeInstruction = "You have run out of investigation steps. Do NOT call any tools. " +
-	"Based solely on what you have already gathered above, write your final answer now in the required format. " +
-	"If you could not confirm some findings, omit them or mark them as unverified, but produce the answer."
-
 func main() {
 	if err := run(); err != nil {
 		fmt.Fprintln(os.Stderr, "gadfly:", err)
@@ -123,6 +113,14 @@ func main() {
 
 func run() error {
 	start := time.Now()
+
+	// Consolidation mode: not a review at all — read the per-model findings the
+	// swarm wrote and print the single cross-model consensus comment. entrypoint.sh
+	// runs this once, after every model has finished.
+	if strings.TrimSpace(os.Getenv("GADFLY_CONSOLIDATE_DIR")) != "" {
+		return runConsolidate()
+	}
+
 	repoDir := os.Getenv("GADFLY_REPO_DIR")
 	diffFile := os.Getenv("GADFLY_DIFF_FILE")
 	systemFile := os.Getenv("GADFLY_SYSTEM_FILE")
@@ -169,7 +167,18 @@ func run() error {
 		} else if worker != nil {
 			fsTools.worker = worker
 		}
-		eng = &majordomoEngine{mdl: mdl, fsTools: fsTools}
+		// The context compactor needs a cheap summarizer; reuse the worker model
+		// when present, else the review model. A bad explicit GADFLY_COMPACT_MODEL
+		// just disables compaction rather than sinking the review.
+		summarizer, serr := resolveSummarizerModel(mdl, fsTools.worker)
+		if serr != nil {
+			fmt.Fprintln(os.Stderr, "gadfly: compaction summarizer disabled:", serr)
+		}
+		rex, rerr := newReviewExecutor(fsTools, mdl, summarizer, os.Getenv("GADFLY_MODEL"), newPRBudget())
+		if rerr != nil {
+			return fmt.Errorf("build review executor: %w", rerr)
+		}
+		eng = &majordomoEngine{rex: rex, mdl: mdl}
 	}
 
 	specialists, registry, auto, serrs := resolveSpecialists(repoDir)
@@ -214,59 +223,64 @@ func run() error {
 	// Optional, best-effort telemetry. OFF unless GADFLY_FINDINGS_URL is set;
 	// any failure is logged to stderr and never affects stdout or the exit code.
 	emit(results, time.Since(start))
+
+	// Optional per-model findings artifact for the cross-model consolidation
+	// pass. No-op unless GADFLY_FINDINGS_OUT is set (entrypoint sets it for a
+	// multi-model swarm). Best-effort, never affects stdout or the exit code.
+	writeFindingsOut(results)
 	return nil
 }
 
 // runSpecialists reviews the diff through each lens and returns the results in
-// the SAME order as specialists, regardless of finish order. Up to
+// the SAME order as specialists, regardless of finish order. It uses executus's
-// GADFLY_LENS_CONCURRENCY lenses run concurrently; the default of 1 keeps the
+// fanout primitive: up to GADFLY_LENS_CONCURRENCY lenses run concurrently (the
-// suite sequential, exactly as before. Each lens already runs under its own
+// default of 1 keeps the suite sequential, exactly as before), and fanout.Run
-// per-lens timeout (reviewWithSpecialist), so concurrency simply overlaps those
+// returns one result per lens in input order. Each lens already runs under its
-// independent passes — and because reviewWithSpecialist builds a fresh toolbox
+// own per-lens timeout (reviewWithSpecialist) and the lenses only read the
-// per pass and the lenses only read the immutable repoFS, they share no mutable
+// immutable repoFS, so concurrency simply overlaps independent passes.
-// state. Results are stored by index so the consolidated comment keeps the
-// configured lens order.
 //
 // Caution: this fans out WITHIN one model. It multiplies with entrypoint.sh's
 // per-provider model concurrency, so total concurrent backend requests ≈
 // (models at once) × (lenses at once). To fan lenses out without oversubscribing
 // the backend, run models one at a time (provider lane cap 1) and raise this.
 func runSpecialists(eng reviewEngine, base string, specialists []Specialist, task, diff string) []specialistResult {
-	results := make([]specialistResult, len(specialists))
-
 	// Optional live status board: publishes this model's per-lens progress to a
 	// file the entrypoint board renders. Inert (no-op) unless GADFLY_STATUS_FILE
 	// is set, so plain runs are unaffected.
 	sw := newStatusWriter(os.Getenv("GADFLY_MODEL"), modelProvider(), specialists)
 
-	conc := min(lensConcurrency(), len(specialists))
+	fanResults := fanout.Run(context.Background(), specialists, fanout.Options[Specialist]{
-
+		MaxConcurrent: lensConcurrency(),
-	sem := make(chan struct{}, conc)
+	}, func(_ context.Context, sp Specialist) (res specialistResult, _ error) {
-	var wg sync.WaitGroup
+		// A panic in one lens must not crash the whole binary (which would kill
-	for i, sp := range specialists {
+		// every other lens's output) or leave this lens stuck at "running" on the
-		wg.Add(1)
+		// status board. fanout does not recover fn panics, so we do it here:
-		sem <- struct{}{} // blocks once `conc` lenses are already in flight
+		// record the panic as an errored result and mark the lens finished.
-		go func(i int, sp Specialist) {
-			defer wg.Done()
-			defer func() { <-sem }()
-			// A panic in one lens must not crash the whole binary (which would
-			// kill every other lens's output) or leave this lens stuck at
-			// "running" on the status board. Recover, record it as an errored
-			// result, and mark the lens finished so the board can complete.
 		defer func() {
 			if r := recover(); r != nil {
-					results[i] = specialistResult{spec: sp, out: fmt.Sprintf("⚠️ This reviewer panicked: %v", r), verdict: verdictUnknown, errored: true}
+				res = specialistResult{spec: sp, out: fmt.Sprintf("⚠️ This reviewer panicked: %v", r), verdict: verdictUnknown, errored: true}
 				sw.set(sp.Name, lensFinished, "", true)
 			}
 		}()
 		sw.set(sp.Name, lensRunning, "", false)
 		out, errored := reviewWithSpecialist(eng, base, sp, task, diff)
 		v := parseVerdict(out)
-			results[i] = specialistResult{spec: sp, out: out, verdict: v, errored: errored}
 		sw.set(sp.Name, lensFinished, v.label(), errored)
-		}(i, sp)
+		return specialistResult{spec: sp, out: out, verdict: v, errored: errored}, nil
+	})
+
+	// fanout guarantees input order; its Result.Err is set only when the context
+	// is cancelled before a lens ran (reviewWithSpecialist embeds its own failures
+	// in the result), so surface that as an errored lens rather than dropping it.
+	results := make([]specialistResult, len(specialists))
+	for i, r := range fanResults {
+		if r.Err != nil {
+			results[i] = specialistResult{spec: specialists[i], out: fmt.Sprintf("⚠️ This reviewer did not run: %v", r.Err), verdict: verdictUnknown, errored: true}
+			sw.set(specialists[i].Name, lensFinished, "", true)
+			continue
+		}
+		results[i] = r.Value
 	}
-	wg.Wait()
 	return results
 }
 
@@ -307,8 +321,7 @@ func providerOverride(envName, provider string) (int, bool) {
 // returned bool is true when the review pass failed (rendered as an inline
 // notice — advisory; one lens failing never sinks the others or the job).
 func reviewWithSpecialist(eng reviewEngine, base string, sp Specialist, task, diff string) (string, bool) {
-	timeout := time.Duration(envInt("GADFLY_TIMEOUT_SECS", defaultTimeoutSecs)) * time.Second
+	ctx, cancel := context.WithTimeout(context.Background(), reviewTimeout())
-	ctx, cancel := context.WithTimeout(context.Background(), timeout)
 	defer cancel()
 
 	draft, err := eng.runPass(ctx, composeSpecialistPrompt(base, sp), task,
@@ -331,90 +344,6 @@ func reviewWithSpecialist(eng reviewEngine, base string, sp Specialist, task, di
 	return final, false
 }
 
-// runAgent runs one agent pass (its own fresh toolbox over the sandbox) and
-// returns the final answer. An empty answer is an error — the caller decides
-// whether that is fatal (review pass) or recoverable (recheck pass). A
-// non-empty answer that ended on a budget/guard error is still returned: the
-// model wrote its output, then ran out of steps.
-//
-// Two mechanisms keep a step-hungry model from hard-failing with no output:
-//  1. A wrap-up steer: once the run comes within wrapUpReserve steps of the
-//     cap, a forceful "stop calling tools, write your final answer" message is
-//     injected so the model spends its remaining steps finalizing.
-//  2. A finalization fallback: if the loop still ends empty (the model ignored
-//     the nudge, or a loop guard tripped), one tool-free model call forces a
-//     final answer out of the transcript already gathered.
-func runAgent(ctx context.Context, mdl llm.Model, fsTools *repoFS, system, task string, maxSteps int) (string, error) {
-	box, err := fsTools.toolbox()
-	if err != nil {
-		return "", err
-	}
-	loop := agent.New(mdl, system,
-		agent.WithToolbox(box),
-		agent.WithMaxSteps(maxSteps),
-		// Guard rails: stop the model from spinning on failing or identical
-		// tool calls instead of writing its answer.
-		agent.WithToolErrorLimits(4, 4),
-	)
-
-	wrapUpAt := maxSteps - wrapUpReserve()
-	if wrapUpAt < 1 {
-		wrapUpAt = 1
-	}
-	var completed int // steps finished so far (updated after each step)
-	nudged := false
-
-	res, runErr := loop.Run(ctx, task,
-		agent.OnStep(func(s agent.Step) { completed = s.Index + 1 }),
-		agent.WithSteer(func() []llm.Message {
-			if !nudged && completed >= wrapUpAt {
-				nudged = true
-				return []llm.Message{llm.UserText(wrapUpInstruction)}
-			}
-			return nil
-		}),
-	)
-
-	out := ""
-	if res != nil {
-		out = strings.TrimSpace(res.Output)
-	}
-	if out != "" {
-		return out, nil
-	}
-
-	// No final answer. If we still have budget on the clock and a transcript to
-	// work from, force a tool-free finalization rather than losing the pass.
-	if res != nil && len(res.Messages) > 0 && ctx.Err() == nil {
-		if forced := forceFinalAnswer(ctx, mdl, system, res.Messages); forced != "" {
-			return forced, nil
-		}
-	}
-
-	if runErr != nil {
-		return "", runErr
-	}
-	return "", errors.New("agent produced no output")
-}
-
-// forceFinalAnswer makes one tool-free model call to squeeze a final answer out
-// of an agent that exhausted its step budget without producing one. Tools are
-// forbidden (ToolChoice "none") so the model must synthesize from the transcript
-// instead of investigating further. Best-effort: any error or empty reply
-// returns "" and the caller falls back to its normal empty-output handling.
-func forceFinalAnswer(ctx context.Context, mdl llm.Model, system string, transcript []llm.Message) string {
-	msgs := append(append([]llm.Message(nil), transcript...), llm.UserText(finalizeInstruction))
-	resp, err := mdl.Generate(ctx, llm.Request{
-		System:     system,
-		Messages:   msgs,
-		ToolChoice: "none",
-	})
-	if err != nil || resp == nil {
-		return ""
-	}
-	return strings.TrimSpace(resp.Text())
-}
-
 // wrapUpReserve is how many steps before the cap the wrap-up nudge fires,
 // overridable via GADFLY_WRAPUP_RESERVE.
 func wrapUpReserve() int {
@@ -431,7 +360,7 @@ func buildTask(diff string) string {
 	truncNote := ""
 	if maxDiff > 0 && len(diff) > maxDiff {
 		diff = diff[:maxDiff]
-		truncNote = fmt.Sprintf("\n\n[NOTE: diff truncated to %d chars in this message; read the changed files (or call get_diff, if available) for the full text.]", maxDiff)
+		truncNote = fmt.Sprintf("\n\n[NOTE: diff truncated to %d chars in this message; page the full diff with get_diff (paginated; pass a `path` to scope it to one file) or read the changed files.]", maxDiff)
 	}
 
 	var b strings.Builder
@@ -458,3 +387,23 @@ func envInt(name string, def int) int {
 	}
 	return n
 }
+
+// envBool reads a boolean-ish env var: def when unset, false for an explicit
+// falsey value (0/false/no/off), true otherwise. The shared spelling for
+// gadfly's "on unless disabled" opt-out flags (GADFLY_RECHECK, GADFLY_COMPACT).
+func envBool(name string, def bool) bool {
+	switch strings.ToLower(strings.TrimSpace(os.Getenv(name))) {
+	case "":
+		return def
+	case "0", "false", "no", "off":
+		return false
+	default:
+		return true
+	}
+}
+
+// reviewTimeout is the per-specialist-lens deadline (GADFLY_TIMEOUT_SECS), shared
+// across a lens's review+recheck passes and applied as each pass's run cap.
+func reviewTimeout() time.Duration {
+	return time.Duration(envInt("GADFLY_TIMEOUT_SECS", defaultTimeoutSecs)) * time.Second
+}

cmd/gadfly/model.go

@@ -127,6 +127,27 @@ func resolveWorkerModel() (llm.Model, error) {
 	return majordomo.Parse(buildSpec(provider, spec))
 }
 
+// resolveSummarizerModel picks the model the context compactor uses to compress
+// the runaway middle of a transcript. It should be CHEAP, since it fires once per
+// compaction: GADFLY_COMPACT_MODEL if set (honoring GADFLY_PROVIDER for a bare
+// id), else the delegate worker model when one is configured (already cheap by
+// design), else the review model itself. Returns (nil, err) only on an explicit
+// bad GADFLY_COMPACT_MODEL spec — the caller logs it and simply runs without
+// compaction rather than failing the review.
+func resolveSummarizerModel(review, worker llm.Model) (llm.Model, error) {
+	if spec := strings.TrimSpace(os.Getenv("GADFLY_COMPACT_MODEL")); spec != "" {
+		provider := strings.TrimSpace(os.Getenv("GADFLY_PROVIDER"))
+		if provider == "" {
+			provider = defaultProvider
+		}
+		return majordomo.Parse(buildSpec(provider, spec))
+	}
+	if worker != nil {
+		return worker, nil
+	}
+	return review, nil
+}
+
 // buildSpec turns (provider, model) into a majordomo spec. A model id that
 // already carries a "provider/" prefix (or is a multi-element failover chain)
 // is passed through verbatim; a bare id is prefixed with the provider.

cmd/gadfly/recheck.go

@@ -2,7 +2,6 @@ package main
 
 import (
 	"fmt"
-	"os"
 	"strings"
 )
 
@@ -11,6 +10,13 @@ import (
 // than discovering them.
 const defaultRecheckMaxSteps = 16
 
+// defaultRecheckDiffChars caps the diff embedded in the recheck task. It is much
+// smaller than the review task's GADFLY_MAX_DIFF_CHARS: the recheck already has
+// the draft findings to verify and can pull the exact hunks it needs via the
+// paginated get_diff tool (optionally scoped to a path), so re-embedding the
+// whole diff on every recheck step is pure burn. Override: GADFLY_RECHECK_DIFF_CHARS.
+const defaultRecheckDiffChars = 20000
+
 // recheckSystemPrompt drives the second, adversarial verification pass. The
 // model is given a DRAFT review and must independently confirm each finding
 // against the real code before letting it survive — the antidote to a
@@ -49,18 +55,16 @@ Output rules:
 - Do NOT invent new findings; this is a verification gate, not a fresh review.
 - Do NOT include meta-commentary about the verification process or which
   findings you dropped — output only the final, corrected review markdown.
+- The draft ends with a fenced ` + "`gadfly-findings`" + ` JSON block. Regenerate it
+  so it lists ONLY the findings that SURVIVED your verification, in the same schema
+  ({"file","line","severity","confidence","title"}; severity one of
+  critical/high/medium/small/trivial, confidence one of high/medium/low). If every
+  finding was dropped, emit an empty array ` + "`[]`" + `. Keep the block last.
 - When done investigating, STOP calling tools and reply with the review.`
 
 // recheckEnabled reports whether the verification pass should run. On unless
 // GADFLY_RECHECK is explicitly a falsey value.
-func recheckEnabled() bool {
+func recheckEnabled() bool { return envBool("GADFLY_RECHECK", true) }
-	switch strings.ToLower(strings.TrimSpace(os.Getenv("GADFLY_RECHECK"))) {
-	case "0", "false", "no", "off":
-		return false
-	default:
-		return true
-	}
-}
 
 // shouldRecheck decides whether to run the verification pass for a given draft.
 // A clean "no material issues" draft has nothing to verify, so it is skipped
@@ -79,11 +83,11 @@ func shouldRecheck(draft string) bool {
 // scrutinize, with the full diff available via get_diff (and embedded here,
 // truncated, to save a tool call).
 func buildRecheckTask(draft, diff string) string {
-	maxDiff := envInt("GADFLY_MAX_DIFF_CHARS", defaultMaxDiffChars)
+	maxDiff := envInt("GADFLY_RECHECK_DIFF_CHARS", defaultRecheckDiffChars)
 	truncNote := ""
 	if maxDiff > 0 && len(diff) > maxDiff {
 		diff = diff[:maxDiff]
-		truncNote = fmt.Sprintf("\n\n[NOTE: diff truncated to %d chars here; read the changed files (or call get_diff, if available) for the full text.]", maxDiff)
+		truncNote = fmt.Sprintf("\n\n[NOTE: diff truncated to %d chars here; call get_diff (paginated; pass a `path` to scope it to one file) or read the changed files for the rest.]", maxDiff)
 	}
 
 	var b strings.Builder

cmd/gadfly/recheck_test.go

@@ -49,7 +49,7 @@ func TestRecheckEnabled(t *testing.T) {
 }
 
 func TestBuildRecheckTask(t *testing.T) {
-	t.Setenv("GADFLY_MAX_DIFF_CHARS", "")
+	t.Setenv("GADFLY_RECHECK_DIFF_CHARS", "")
 	draft := "VERDICT: Blocking issues found\n- foo.go:1 broken"
 	out := buildRecheckTask(draft, "diff --git a/x b/x\n+y\n")
 	if !strings.Contains(out, draft) {
@@ -77,25 +77,46 @@ func fakeModel(t *testing.T, reply string) llm.Model {
 	return m
 }
 
-func TestRunAgent_ReturnsOutput(t *testing.T) {
+// newTestReviewExecutor builds a reviewExecutor over a fake model + repo for unit
+// tests: no compaction summarizer and no budget, so it exercises the bare agent
+// loop through the executus run kernel without any network call.
+func newTestReviewExecutor(t *testing.T, mdl llm.Model, fs *repoFS) *reviewExecutor {
+	t.Helper()
+	rex, err := newReviewExecutor(fs, mdl, nil, "mock", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return rex
+}
+
+// testEngine wraps newTestReviewExecutor in a majordomoEngine for the
+// runSpecialists-level tests.
+func testEngine(t *testing.T, mdl llm.Model, fs *repoFS) *majordomoEngine {
+	t.Helper()
+	return &majordomoEngine{rex: newTestReviewExecutor(t, mdl, fs), mdl: mdl}
+}
+
+func TestReviewExecutor_ReturnsOutput(t *testing.T) {
 	fs, err := newRepoFS(t.TempDir(), "diff")
 	if err != nil {
 		t.Fatal(err)
 	}
 	mdl := fakeModel(t, "  corrected review: No material issues found.  ")
-	out, err := runAgent(context.Background(), mdl, fs, "sys", "task", 4)
+	rex := newTestReviewExecutor(t, mdl, fs)
+	out, err := rex.run(context.Background(), "sys", "task", 4)
 	if err != nil {
-		t.Fatalf("runAgent: %v", err)
+		t.Fatalf("run: %v", err)
 	}
 	if out != "corrected review: No material issues found." {
-		t.Errorf("runAgent should return trimmed model output, got %q", out)
+		t.Errorf("run should return trimmed model output, got %q", out)
 	}
 }
 
-func TestRunAgent_EmptyIsError(t *testing.T) {
+func TestReviewExecutor_EmptyIsError(t *testing.T) {
 	fs, _ := newRepoFS(t.TempDir(), "diff")
 	mdl := fakeModel(t, "   ")
-	if _, err := runAgent(context.Background(), mdl, fs, "sys", "task", 4); err == nil {
+	rex := newTestReviewExecutor(t, mdl, fs)
-		t.Error("runAgent should error on empty model output")
+	if _, err := rex.run(context.Background(), "sys", "task", 4); err == nil {
+		t.Error("run should error on empty model output")
 	}
 }

cmd/gadfly/review.go

@@ -0,0 +1,296 @@
+package main
+
+// Inline PR review. After the consensus comment is rendered, Gadfly also posts a
+// single Gitea pull review (state COMMENT — advisory, NEVER request-changes or
+// approve) whose inline comments anchor consensus findings to the exact changed
+// lines. The issue comment stays the ranked overview; the review puts each
+// finding next to the code it's about — the "reviewer integrated with Gitea" the
+// project wanted, without ever blocking a merge.
+//
+// All of this is best-effort: disabled by GADFLY_INLINE_REVIEW=0 or when the diff
+// / API creds aren't available, only anchors findings that land on a line present
+// in the diff (Gitea rejects comments off the diff), and any error is logged to
+// stderr without touching the consensus comment (already on stdout) or the exit
+// code.
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"regexp"
+	"strconv"
+	"strings"
+	"time"
+)
+
+const (
+	// inlineReviewMarker tags our review body so a re-run can delete the previous
+	// one instead of stacking duplicate inline comments.
+	inlineReviewMarker = "<!-- gadfly-inline-review -->"
+	// maxInlineComments caps how many inline comments one review carries, so a
+	// huge diff can't produce a wall of annotations. Clusters are pre-sorted by
+	// agreement×severity, so the cap keeps the most important.
+	maxInlineComments       = 25
+	inlineReviewHTTPTimeout = 20 * time.Second
+)
+
+// reviewComment is one inline comment in a Gitea pull review. Field names match
+// the Gitea API EXACTLY (new_position = line in the new/head file).
+type reviewComment struct {
+	Path        string `json:"path"`
+	Body        string `json:"body"`
+	NewPosition int    `json:"new_position"`
+}
+
+// createReview is the POST /pulls/{n}/reviews body. event is ALWAYS "COMMENT".
+type createReview struct {
+	Body     string          `json:"body"`
+	Event    string          `json:"event"`
+	Comments []reviewComment `json:"comments"`
+}
+
+// postInlineReview posts one COMMENT-state pull review with inline comments for
+// consensus findings on changed lines. No-op + best-effort (see file comment).
+func postInlineReview(clusters []cluster) {
+	if strings.EqualFold(strings.TrimSpace(os.Getenv("GADFLY_INLINE_REVIEW")), "0") {
+		return
+	}
+	api := strings.TrimRight(strings.TrimSpace(os.Getenv("GITEA_API")), "/")
+	token := strings.TrimSpace(os.Getenv("GITEA_TOKEN"))
+	pr := strings.TrimSpace(os.Getenv("GADFLY_PR"))
+	diffPath := strings.TrimSpace(os.Getenv("GADFLY_DIFF_FILE"))
+	if api == "" || token == "" || pr == "" || diffPath == "" {
+		return
+	}
+	diff, err := os.ReadFile(diffPath)
+	if err != nil {
+		fmt.Fprintln(os.Stderr, "gadfly: inline review: read diff:", err)
+		return
+	}
+
+	comments := inlineComments(clusters, parseDiffNewLines(string(diff)))
+	if len(comments) == 0 {
+		return // nothing anchors to a changed line; the consensus comment covers it
+	}
+
+	client := &http.Client{Timeout: inlineReviewHTTPTimeout}
+	base := fmt.Sprintf("%s/pulls/%s/reviews", api, pr)
+	deletePriorReviews(client, base, token) // avoid stacking on re-runs
+
+	body := fmt.Sprintf("%s\n🪰 **Gadfly consensus review** — %d inline finding%s on changed lines. See the consensus comment for the full ranked summary.\n\n<sub>Advisory only — does not block merge.</sub>",
+		inlineReviewMarker, len(comments), plural(len(comments)))
+	if err := giteaSend(client, http.MethodPost, base, token, createReview{Body: body, Event: "COMMENT", Comments: comments}); err != nil {
+		fmt.Fprintln(os.Stderr, "gadfly: inline review post:", err)
+	}
+}
+
+// inlineComments builds inline comments for the clusters that anchor to a line
+// present in the diff, in priority order (clusters are pre-sorted), capped.
+func inlineComments(clusters []cluster, addable map[string]map[int]bool) []reviewComment {
+	var out []reviewComment
+	for _, c := range clusters {
+		path := normPath(c.file)
+		anchor := anchorLine(addable[path], c.line, c.maxLine)
+		if anchor == 0 {
+			continue
+		}
+		out = append(out, reviewComment{Path: path, NewPosition: anchor, Body: inlineBody(c)})
+		if len(out) >= maxInlineComments {
+			break
+		}
+	}
+	return out
+}
+
+// anchorLine returns the first line in [lo,hi] that is an added line in the diff,
+// or 0 if none. Scanning the cluster's whole span (not just its representative
+// line) anchors a finding whose min line is just outside the diff but whose span
+// still overlaps a changed line.
+func anchorLine(added map[int]bool, lo, hi int) int {
+	if added == nil || lo <= 0 {
+		return 0
+	}
+	if hi < lo { // single-line cluster (maxLine unset)
+		hi = lo
+	}
+	for ln := lo; ln <= hi; ln++ {
+		if added[ln] {
+			return ln
+		}
+	}
+	return 0
+}
+
+// inlineBody renders one inline comment: severity + title, who flagged it, detail.
+func inlineBody(c cluster) string {
+	var b strings.Builder
+	fmt.Fprintf(&b, "%s **%s**", sevIcon(c.severity), strings.TrimSpace(c.title))
+	fmt.Fprintf(&b, "\n\n_%s · flagged by %d model%s_", lensList(c.lenses), len(c.models), plural(len(c.models)))
+	if d := strings.TrimSpace(c.detail); d != "" {
+		fmt.Fprintf(&b, "\n\n%s", d)
+	}
+	b.WriteString("\n\n<sub>🪰 Gadfly · advisory</sub>")
+	return b.String()
+}
+
+// parseDiffNewLines returns, per file, the set of NEW-file line numbers that were
+// ADDED in the unified diff — the safest lines for an inline comment to anchor to
+// (Gitea reliably accepts comments on added lines). Context lines are walked to
+// keep the line counter correct but are NOT recorded: anchoring only to added
+// lines avoids the all-or-nothing review POST being rejected for an off-change
+// anchor. Hunk lengths from the @@ header bound each hunk, so a content line that
+// happens to start with "+++ " or "@@" is still read as content, not a header.
+func parseDiffNewLines(diff string) map[string]map[int]bool {
+	out := map[string]map[int]bool{}
+	var file string
+	var newLine, oldRem, newRem int
+	inHunk := false
+	for _, line := range strings.Split(diff, "\n") {
+		if inHunk && (newRem > 0 || oldRem > 0) {
+			switch {
+			case strings.HasPrefix(line, "+"):
+				record(out, file, newLine) // added line — anchorable
+				newLine++
+				newRem--
+			case strings.HasPrefix(line, "-"):
+				oldRem--
+			case strings.HasPrefix(line, "\\"): // "\ No newline at end of file"
+			default: // context line (leading space, or an empty line): advance, don't record
+				newLine++
+				newRem--
+				oldRem--
+			}
+			if newRem <= 0 && oldRem <= 0 {
+				inHunk = false
+			}
+			continue
+		}
+		switch {
+		case strings.HasPrefix(line, "+++ "):
+			file = normPath(strings.TrimPrefix(line, "+++ "))
+			if file == "/dev/null" {
+				file = ""
+			}
+		case strings.HasPrefix(line, "@@"):
+			if m := hunkRe.FindStringSubmatch(line); m != nil && file != "" {
+				newLine, _ = strconv.Atoi(m[3])
+				oldRem = atoiOr(m[2], 1)
+				newRem = atoiOr(m[4], 1)
+				inHunk = newRem > 0 || oldRem > 0
+			}
+		}
+	}
+	return out
+}
+
+// hunkRe captures a unified-diff hunk header's old/new start+length:
+// @@ -<oldStart>[,<oldLen>] +<newStart>[,<newLen>] @@
+var hunkRe = regexp.MustCompile(`^@@ -(\d+)(?:,(\d+))? \+(\d+)(?:,(\d+))? @@`)
+
+// normPath trims a unified-diff path down to a repo-relative one: strip a single
+// leading "a/" or "b/" prefix (and any "./"), and surrounding whitespace. Applied
+// to BOTH the diff paths and a finding's file so they match even when a model
+// writes "./pkg/x.go" or the diff carries the "b/" prefix.
+func normPath(p string) string {
+	p = strings.TrimSpace(p)
+	p = strings.TrimPrefix(p, "./")
+	if strings.HasPrefix(p, "a/") || strings.HasPrefix(p, "b/") {
+		p = p[2:]
+	}
+	return p
+}
+
+func record(out map[string]map[int]bool, file string, line int) {
+	if file == "" {
+		return
+	}
+	if out[file] == nil {
+		out[file] = map[int]bool{}
+	}
+	out[file][line] = true
+}
+
+// atoiOr parses s, returning def when s is empty or unparseable. Used for the
+// optional hunk-length fields (absent => length 1).
+func atoiOr(s string, def int) int {
+	if s == "" {
+		return def
+	}
+	if n, err := strconv.Atoi(s); err == nil {
+		return n
+	}
+	return def
+}
+
+// deletePriorReviews removes our previous inline reviews (matched by the body
+// marker) so a re-run replaces rather than stacks. Best-effort and quiet.
+func deletePriorReviews(client *http.Client, base, token string) {
+	const perPage = 50
+	for page := 1; page <= 10; page++ { // bound the scan, but page past 50 so a stale marked review isn't missed
+		req, err := http.NewRequest(http.MethodGet, fmt.Sprintf("%s?limit=%d&page=%d", base, perPage, page), nil)
+		if err != nil {
+			return
+		}
+		req.Header.Set("Authorization", "token "+token)
+		resp, err := client.Do(req)
+		if err != nil {
+			return
+		}
+		if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+			io.Copy(io.Discard, resp.Body)
+			resp.Body.Close()
+			return
+		}
+		var reviews []struct {
+			ID   int    `json:"id"`
+			Body string `json:"body"`
+		}
+		_ = json.NewDecoder(resp.Body).Decode(&reviews)
+		resp.Body.Close()
+		for _, r := range reviews {
+			if !strings.Contains(r.Body, inlineReviewMarker) {
+				continue
+			}
+			dreq, err := http.NewRequest(http.MethodDelete, fmt.Sprintf("%s/%d", base, r.ID), nil)
+			if err != nil {
+				continue
+			}
+			dreq.Header.Set("Authorization", "token "+token)
+			if dresp, err := client.Do(dreq); err == nil {
+				io.Copy(io.Discard, dresp.Body)
+				dresp.Body.Close()
+			}
+		}
+		if len(reviews) < perPage {
+			return // last page
+		}
+	}
+}
+
+// giteaSend marshals payload and sends it with Gitea's "token" auth scheme,
+// treating a non-2xx response as an error (with a snippet of the body).
+func giteaSend(client *http.Client, method, url, token string, payload any) error {
+	body, err := json.Marshal(payload)
+	if err != nil {
+		return err
+	}
+	req, err := http.NewRequest(method, url, bytes.NewReader(body))
+	if err != nil {
+		return err
+	}
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Authorization", "token "+token)
+	resp, err := client.Do(req)
+	if err != nil {
+		return err
+	}
+	defer resp.Body.Close()
+	snippet, _ := io.ReadAll(io.LimitReader(resp.Body, 2048))
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return fmt.Errorf("%s %s: status %d: %s", method, url, resp.StatusCode, strings.TrimSpace(string(snippet)))
+	}
+	return nil
+}

cmd/gadfly/review_test.go

@@ -0,0 +1,107 @@
+package main
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestParseDiffNewLines(t *testing.T) {
+	diff := "diff --git a/a.go b/a.go\n" +
+		"index 111..222 100644\n" +
+		"--- a/a.go\n" +
+		"+++ b/a.go\n" +
+		"@@ -1,3 +1,4 @@\n" +
+		" line1\n" +
+		"-old2\n" +
+		"+new2\n" +
+		"+new3\n" +
+		" line4\n"
+	got := parseDiffNewLines(diff)
+	// Only ADDED lines anchor: new2 (line 2) and new3 (line 3). Context lines 1
+	// and 4 are walked for counting but not recorded.
+	want := map[int]bool{2: true, 3: true}
+	if len(got["a.go"]) != len(want) {
+		t.Fatalf("a.go lines = %v, want %v (added-only)", got["a.go"], want)
+	}
+	for ln := range want {
+		if !got["a.go"][ln] {
+			t.Errorf("expected a.go line %d anchorable", ln)
+		}
+	}
+	if got["a.go"][1] || got["a.go"][4] {
+		t.Error("context lines 1/4 should not be anchorable (added-only)")
+	}
+}
+
+func TestParseDiffNewLinesContentLooksLikeHeader(t *testing.T) {
+	// An added line whose CONTENT is "++ weird" appears as "+++ weird" in the
+	// diff. Hunk-length tracking must read it as content, not a file header.
+	diff := "--- a/x\n+++ b/x\n@@ -1,1 +1,2 @@\n ctx\n+++ weird\n"
+	got := parseDiffNewLines(diff)
+	if got["x"][1] { // context line, not recorded (added-only)
+		t.Errorf("line 1 is context, should not anchor: %v", got["x"])
+	}
+	if !got["x"][2] { // the "+++ weird" added line
+		t.Errorf("want added line 2 anchorable, got %v", got["x"])
+	}
+}
+
+func TestAnchorLineScansSpan(t *testing.T) {
+	// A cluster spanning 10..14 whose min line (10) isn't in the diff but whose
+	// span includes added line 14 must anchor to 14, not be dropped.
+	added := map[string]map[int]bool{"a.go": {14: true}}
+	clusters := []cluster{{file: "a.go", line: 10, maxLine: 14, severity: "high", title: "t", models: set("m1"), lenses: set("x")}}
+	cs := inlineComments(clusters, added)
+	if len(cs) != 1 || cs[0].NewPosition != 14 {
+		t.Fatalf("want anchor at 14 via span scan, got %+v", cs)
+	}
+}
+
+func TestParseDiffNewLinesMultiFile(t *testing.T) {
+	diff := "diff --git a/one.go b/one.go\n--- a/one.go\n+++ b/one.go\n@@ -0,0 +1,2 @@\n+a\n+b\n" +
+		"diff --git a/two.go b/two.go\n--- a/two.go\n+++ b/two.go\n@@ -5,0 +6,1 @@\n+c\n"
+	got := parseDiffNewLines(diff)
+	if !got["one.go"][1] || !got["one.go"][2] {
+		t.Errorf("one.go: want 1,2 got %v", got["one.go"])
+	}
+	if !got["two.go"][6] {
+		t.Errorf("two.go: want 6 got %v", got["two.go"])
+	}
+}
+
+func TestInlineCommentsNormalizesPaths(t *testing.T) {
+	// Diff path "b/pkg/x.go" -> "pkg/x.go"; a finding written as "./pkg/x.go" must
+	// still anchor (both normalize to the same repo-relative path).
+	addable := parseDiffNewLines("--- a/pkg/x.go\n+++ b/pkg/x.go\n@@ -0,0 +1,1 @@\n+bug\n")
+	clusters := []cluster{{file: "./pkg/x.go", line: 1, severity: "high", title: "t", models: set("m1"), lenses: set("x")}}
+	if cs := inlineComments(clusters, addable); len(cs) != 1 || cs[0].Path != "pkg/x.go" {
+		t.Errorf("path normalization failed to anchor: %+v", cs)
+	}
+}
+
+func TestInlineCommentsFiltersToDiffLines(t *testing.T) {
+	addable := map[string]map[int]bool{"a.go": {10: true, 11: true}}
+	clusters := []cluster{
+		{file: "a.go", line: 10, severity: "high", title: "anchored", models: set("m1", "m2"), lenses: set("security"), detail: "d"},
+		{file: "a.go", line: 99, severity: "high", title: "off-diff line", models: set("m1"), lenses: set("security")},
+		{file: "b.go", line: 1, severity: "high", title: "off-diff file", models: set("m1"), lenses: set("security")},
+	}
+	cs := inlineComments(clusters, addable)
+	if len(cs) != 1 {
+		t.Fatalf("want 1 anchorable inline comment, got %d", len(cs))
+	}
+	if cs[0].Path != "a.go" || cs[0].NewPosition != 10 {
+		t.Errorf("wrong anchor: %+v", cs[0])
+	}
+	if !strings.Contains(cs[0].Body, "anchored") || !strings.Contains(cs[0].Body, "2 model") {
+		t.Errorf("inline body missing title/agreement: %q", cs[0].Body)
+	}
+}
+
+func set(xs ...string) map[string]bool {
+	m := map[string]bool{}
+	for _, x := range xs {
+		m[x] = true
+	}
+	return m
+}

cmd/gadfly/tools.go

@@ -9,6 +9,7 @@ import (
 	"regexp"
 	"sort"
 	"strings"
+	"sync"
 
 	llm "gitea.stevedudenhoeffer.com/steve/majordomo/llm"
 )
@@ -22,6 +23,8 @@ const (
 	maxGrepResults  = 200       // per grep call
 	maxFindResults  = 200       // per find_files call
 	maxLineLen      = 400       // truncate any single returned line to this
+	maxGetDiffLines = 800       // per get_diff call (paginated window)
+	maxGetDiffBytes = 64 * 1024 // per get_diff call
 )
 
 // skipDirs are never descended into by grep / find_files — noise and bulk that
@@ -40,6 +43,12 @@ type repoFS struct {
 	root   string    // absolute, symlink-resolved repo root
 	diff   string    // the full PR unified diff (served by get_diff)
 	worker llm.Model // optional cheap model for delegate_investigation; nil = no delegation
+
+	// diffLines caches the split diff so paging through get_diff doesn't re-split
+	// the whole (possibly large) diff on every call. The diff is immutable, so the
+	// cache is computed once and safe to share across concurrent lenses.
+	diffOnce  sync.Once
+	diffLines []string
 }
 
 // newRepoFS resolves root to an absolute, symlink-free path.
@@ -108,15 +117,24 @@ func (r *repoFS) fsTools() []llm.Tool {
 	}
 }
 
-// toolbox builds the reviewer's toolbox: the read-only repo tools, plus the
+// allTools is the single source of truth for the reviewer's tool set: the
-// delegate_investigation tool when a worker model is configured.
+// read-only repo tools, plus delegate_investigation when a worker model is
-func (r *repoFS) toolbox() (*llm.Toolbox, error) {
+// configured. Both the executus registry (gadflyToolRegistry, the production
-	box := llm.NewToolbox("gadfly")
+// path) and toolbox() build from this list.
+func (r *repoFS) allTools() []llm.Tool {
 	tools := r.fsTools()
 	if r.worker != nil {
 		tools = append(tools, r.delegateTool())
 	}
-	for _, t := range tools {
+	return tools
+}
+
+// toolbox builds a majordomo toolbox from allTools(). The production review path
+// now goes through executus's tool.Registry (see executus.go); this remains for
+// the toolbox-level tests (the `call` helper).
+func (r *repoFS) toolbox() (*llm.Toolbox, error) {
+	box := llm.NewToolbox("gadfly")
+	for _, t := range r.allTools() {
 		if err := box.Add(t); err != nil {
 			return nil, fmt.Errorf("add tool %q: %w", t.Name, err)
 		}
@@ -396,15 +414,124 @@ func (r *repoFS) findFilesTool() llm.Tool {
 	)
 }
 
+type getDiffArgs struct {
+	Path      string `json:"path,omitempty" description:"Optional changed-file path (e.g. pkg/foo/bar.go); returns ONLY that file's diff hunks. Omit for the whole diff. Use this on a large PR to pull just the file a finding is about."`
+	StartLine int    `json:"start_line,omitempty" description:"Optional 1-based line to start from within the (whole or path-scoped) diff (default 1)."`
+	Limit     int    `json:"limit,omitempty" description:"Optional max number of diff lines to return (default/maximum 800)."`
+}
+
 func (r *repoFS) getDiffTool() llm.Tool {
-	return llm.DefineTool[struct{}](
+	return llm.DefineTool[getDiffArgs](
 		"get_diff",
-		"Return the complete unified diff under review. The diff is also included (possibly truncated) in the task message; call this to get the full, untruncated text.",
+		"Return the unified diff under review as a numbered, PAGINATED window (like read_file) — not the whole diff at once, so a huge PR can't blow the context window. Pass `path` to fetch just one changed file's hunks, or `start_line`/`limit` to page through. A truncated copy of the diff is also embedded in the task message.",
-		func(_ context.Context, _ struct{}) (any, error) {
+		func(_ context.Context, args getDiffArgs) (any, error) {
-			if strings.TrimSpace(r.diff) == "" {
+			scope := "the diff"
+			var lines []string
+			if p := strings.TrimSpace(args.Path); p != "" {
+				lines = diffLinesForPath(r.diff, p)
+				if len(lines) == 0 {
+					return fmt.Sprintf("(no diff hunks for path %q; check it against the changed-files list — the path must match a file the diff touches)", p), nil
+				}
+				scope = "the diff for " + p
+			} else {
+				lines = r.diffAllLines()
+				if len(lines) == 0 {
 					return "(empty diff)", nil
 				}
-			return r.diff, nil
+			}
+			return windowDiff(lines, scope, args.StartLine, args.Limit), nil
 		},
 	)
 }
+
+// diffAllLines returns the whole diff split into lines, cached so paging never
+// re-splits the (possibly large) diff.
+func (r *repoFS) diffAllLines() []string {
+	r.diffOnce.Do(func() { r.diffLines = splitDiffLines(r.diff) })
+	return r.diffLines
+}
+
+// splitDiffLines splits a unified diff into lines, dropping the single trailing
+// empty element a trailing newline produces — otherwise windowDiff would emit a
+// blank final line and over-count the total by one.
+func splitDiffLines(diff string) []string {
+	lines := strings.Split(diff, "\n")
+	if n := len(lines); n > 0 && lines[n-1] == "" {
+		lines = lines[:n-1]
+	}
+	return lines
+}
+
+// windowDiff returns a numbered, paginated slice of pre-split diff lines,
+// mirroring read_file's caps (maxGetDiffLines / maxGetDiffBytes / maxLineLen) so
+// a single get_diff call can never dump a multi-hundred-KB diff into the
+// transcript — the amplifier behind the large-PR token burn. The full diff stays
+// reachable by paging with start_line, or scoped per file via the path arg.
+func windowDiff(lines []string, scope string, start, limit int) string {
+	total := len(lines)
+	if start < 1 {
+		start = 1
+	}
+	if limit <= 0 || limit > maxGetDiffLines {
+		limit = maxGetDiffLines
+	}
+	if start > total {
+		return fmt.Sprintf("(%s has %d lines; nothing at/after line %d)", scope, total, start)
+	}
+	var b strings.Builder
+	emitted := 0
+	i := start - 1
+	for ; i < total; i++ {
+		if emitted >= limit || b.Len() >= maxGetDiffBytes {
+			break
+		}
+		line := lines[i]
+		if len(line) > maxLineLen {
+			line = line[:maxLineLen] + "…"
+		}
+		fmt.Fprintf(&b, "%d\t%s\n", i+1, line)
+		emitted++
+	}
+	if i < total {
+		// i (0-based) is the first line NOT emitted; line i was the last shown.
+		fmt.Fprintf(&b, "... (%s truncated after line %d of %d; call get_diff again with start_line=%d for the rest, or pass a `path` to scope to one file)\n", scope, i, total, i+1)
+	}
+	return b.String()
+}
+
+// diffLinesForPath returns the unified-diff lines for one changed file: from the
+// `diff --git` header that names path through to the next header (or end). The
+// header names two path tokens (a/<old> b/<new>); a match is on a WHOLE token,
+// so path "foo.go" does not pull in "barfoo.go" — and a trailing "/" scopes to a
+// directory (e.g. "pkg/foo/" matches pkg/foo/bar.go).
+func diffLinesForPath(diff, path string) []string {
+	want := strings.TrimPrefix(strings.TrimPrefix(strings.TrimSpace(path), "a/"), "b/")
+	var out []string
+	inSection := false
+	for _, ln := range splitDiffLines(diff) {
+		if strings.HasPrefix(ln, "diff --git ") {
+			inSection = diffHeaderNames(ln, want)
+		}
+		if inSection {
+			out = append(out, ln)
+		}
+	}
+	return out
+}
+
+// diffHeaderNames reports whether a `diff --git a/X b/Y` header names want as one
+// of its (a/b-stripped) path tokens — exact whole-token match, or a directory
+// prefix when want ends with "/".
+func diffHeaderNames(header, want string) bool {
+	fields := strings.Fields(header)
+	if len(fields) < 3 {
+		return false
+	}
+	for _, f := range fields[2:] {
+		p := strings.TrimPrefix(strings.TrimPrefix(f, "a/"), "b/")
+		if p == want || (strings.HasSuffix(want, "/") && strings.HasPrefix(p, want)) {
+			return true
+		}
+	}
+	return false
+}

cmd/gadfly/tools_test.go

@@ -197,15 +197,93 @@ func TestFindFilesTool(t *testing.T) {
 
 func TestGetDiffTool(t *testing.T) {
 	root := buildFixtureRepo(t)
-	const diff = "diff --git a/x b/x\n+added line\n"
+	const diff = "diff --git a/x b/x\n--- a/x\n+++ b/x\n+added line\n" +
+		"diff --git a/y.go b/y.go\n--- a/y.go\n+++ b/y.go\n+y change\n"
 	fs, _ := newRepoFS(root, diff)
 
+	// Default: the whole diff as a NUMBERED window (paginated), not a raw dump —
+	// so a huge PR can't be poured into the transcript in one call.
+	out, err := call(t, fs, "get_diff", map[string]any{})
+	if err != nil {
+		t.Fatalf("get_diff: %v", err)
+	}
+	if !strings.Contains(out, "1\tdiff --git a/x b/x") {
+		t.Errorf("get_diff should return a numbered window, got:\n%s", out)
+	}
+	if !strings.Contains(out, "+added line") || !strings.Contains(out, "+y change") {
+		t.Errorf("the full (short) diff window should include every hunk, got:\n%s", out)
+	}
+
+	// path filter: only the named file's hunks come back.
+	out, err = call(t, fs, "get_diff", map[string]any{"path": "y.go"})
+	if err != nil {
+		t.Fatalf("get_diff path: %v", err)
+	}
+	if !strings.Contains(out, "y change") {
+		t.Errorf("get_diff path=y.go should include y's hunk, got:\n%s", out)
+	}
+	if strings.Contains(out, "added line") {
+		t.Errorf("get_diff path=y.go must NOT include x's hunk, got:\n%s", out)
+	}
+
+	// unknown path: a clear note, never an error.
+	out, err = call(t, fs, "get_diff", map[string]any{"path": "nope.txt"})
+	if err != nil {
+		t.Fatalf("get_diff unknown path: %v", err)
+	}
+	if !strings.Contains(out, "no diff hunks") {
+		t.Errorf("get_diff for an unknown path should note no hunks, got:\n%s", out)
+	}
+}
+
+// TestGetDiffTool_Paginates: a diff longer than the per-call line cap is returned
+// as a truncated window with a paging hint, and start_line pages past it — the
+// mechanism that stops get_diff from dumping a multi-hundred-KB diff at once.
+func TestGetDiffTool_Paginates(t *testing.T) {
+	diff := "diff --git a/big b/big\n" + strings.Repeat("+line\n", maxGetDiffLines+50)
+	fs, _ := newRepoFS(t.TempDir(), diff)
+
 	out, err := call(t, fs, "get_diff", map[string]any{})
 	if err != nil {
 		t.Fatalf("get_diff: %v", err)
 	}
-	if out != diff {
+	if !strings.Contains(out, "truncated after line") {
-		t.Errorf("get_diff returned %q, want %q", out, diff)
+		t.Error("a diff longer than the per-call cap should be truncated with a paging hint")
+	}
+	if strings.Contains(out, "801\t") {
+		t.Error("the first window must stop at the line cap, not reach line 801")
+	}
+
+	out, err = call(t, fs, "get_diff", map[string]any{"start_line": 805})
+	if err != nil {
+		t.Fatalf("get_diff page: %v", err)
+	}
+	if !strings.Contains(out, "805\t") {
+		t.Error("paging with start_line=805 should include line 805")
+	}
+}
+
+// TestDiffLinesForPath_Anchored: get_diff path= matches on a WHOLE path token,
+// so "foo.go" never pulls in "barfoo.go" (the unanchored-substring weakness the
+// swarm flagged), while a trailing "/" still scopes to a directory.
+func TestDiffLinesForPath_Anchored(t *testing.T) {
+	diff := "diff --git a/foo.go b/foo.go\n+foo change\n" +
+		"diff --git a/barfoo.go b/barfoo.go\n+barfoo change\n" +
+		"diff --git a/pkg/x.go b/pkg/x.go\n+x change\n"
+
+	joined := strings.Join(diffLinesForPath(diff, "foo.go"), "\n")
+	if !strings.Contains(joined, "foo change") {
+		t.Errorf("foo.go should match its own hunk:\n%s", joined)
+	}
+	if strings.Contains(joined, "barfoo change") {
+		t.Errorf("foo.go must NOT match barfoo.go (unanchored substring regression):\n%s", joined)
+	}
+
+	if !strings.Contains(strings.Join(diffLinesForPath(diff, "pkg/"), "\n"), "x change") {
+		t.Error("a trailing-slash path should scope to the directory (pkg/ -> pkg/x.go)")
+	}
+	if len(diffLinesForPath(diff, "nope.go")) != 0 {
+		t.Error("an unknown path should yield no lines")
 	}
 }
 

cmd/gadfly/wrapup_test.go

@@ -37,10 +37,11 @@ func lastUserText(req llm.Request) string {
 	return req.Messages[len(req.Messages)-1].Text()
 }
 
-// TestRunAgent_WrapUpNudgeProducesAnswer: a model that keeps calling tools until
+// TestReviewExecutor_WrapUpNudgeProducesAnswer: a model that keeps calling tools
-// it is nudged to wrap up should still finish inside its budget — the steer
+// until it is nudged to wrap up should still finish inside its budget — the steer
-// message arrives a few steps before the cap and the model writes its answer.
+// message (delivered by the executus wrap-up critic a few steps before the cap)
-func TestRunAgent_WrapUpNudgeProducesAnswer(t *testing.T) {
+// arrives and the model writes its answer.
+func TestReviewExecutor_WrapUpNudgeProducesAnswer(t *testing.T) {
 	t.Setenv("GADFLY_WRAPUP_RESERVE", "4")
 
 	final := "VERDICT: No material issues found."
@@ -60,9 +61,10 @@ func TestRunAgent_WrapUpNudgeProducesAnswer(t *testing.T) {
 	}
 	fs, _ := newRepoFS(t.TempDir(), "diff --git a/x b/x\n+y\n")
 
-	out, err := runAgent(context.Background(), mdl, fs, "sys", "task", 12)
+	rex := newTestReviewExecutor(t, mdl, fs)
+	out, err := rex.run(context.Background(), "sys", "task", 12)
 	if err != nil {
-		t.Fatalf("runAgent should succeed via wrap-up nudge, got error: %v", err)
+		t.Fatalf("run should succeed via wrap-up nudge, got error: %v", err)
 	}
 	if out != final {
 		t.Errorf("expected final review %q, got %q", final, out)
@@ -72,24 +74,19 @@ func TestRunAgent_WrapUpNudgeProducesAnswer(t *testing.T) {
 	}
 }
 
-// TestRunAgent_FinalizationFallback: a model that ignores the wrap-up nudge and
+// TestReviewExecutor_ExhaustionWithoutAnswerIsError: a model that ignores the
-// spins on tools until the cap should NOT hard-fail — the tool-free finalization
+// wrap-up nudge and spins on tools until the step cap produces no final answer.
-// pass forces a final answer out of the transcript.
+// The transcript-based forced-finalization fallback was removed in the executus
-func TestRunAgent_FinalizationFallback(t *testing.T) {
+// re-platform (run.Result does not expose the loop transcript), so the pass now
+// surfaces an error — which reviewWithSpecialist renders as an advisory "reviewer
+// failed to complete" notice rather than a phantom success.
+func TestReviewExecutor_ExhaustionWithoutAnswerIsError(t *testing.T) {
 	t.Setenv("GADFLY_WRAPUP_RESERVE", "2")
 
-	final := "VERDICT: Minor issues\n- something"
-	forcedCalled := false
 	n := 0
-	p := fake.New("fake", fake.WithDefault(func(_ string, req llm.Request) fake.Step {
+	p := fake.New("fake", fake.WithDefault(func(_ string, _ llm.Request) fake.Step {
-		// Only the tool-free finalization pass forbids tools — reply there.
-		if req.ToolChoice == "none" {
-			forcedCalled = true
-			return fake.Reply(final)
-		}
-		// Otherwise keep spinning, ignoring the wrap-up nudge entirely.
 		n++
-		return spinToolCall(n)
+		return spinToolCall(n) // spin forever, ignoring the wrap-up nudge
 	}))
 	mdl, err := p.Model("mock")
 	if err != nil {
@@ -97,37 +94,9 @@ func TestRunAgent_FinalizationFallback(t *testing.T) {
 	}
 	fs, _ := newRepoFS(t.TempDir(), "diff --git a/x b/x\n+y\n")
 
-	out, err := runAgent(context.Background(), mdl, fs, "sys", "task", 6)
+	rex := newTestReviewExecutor(t, mdl, fs)
-	if err != nil {
+	if _, err := rex.run(context.Background(), "sys", "task", 6); err == nil {
-		t.Fatalf("runAgent should recover via finalization fallback, got error: %v", err)
+		t.Error("run should error when the model exhausts its steps without an answer")
-	}
-	if !forcedCalled {
-		t.Error("finalization fallback was never invoked")
-	}
-	if out != final {
-		t.Errorf("expected forced final answer %q, got %q", final, out)
-	}
-}
-
-// TestRunAgent_FallbackStillEmptyIsError: if even the tool-free finalization
-// yields nothing, runAgent surfaces an error rather than a phantom success.
-func TestRunAgent_FallbackStillEmptyIsError(t *testing.T) {
-	n := 0
-	p := fake.New("fake", fake.WithDefault(func(_ string, req llm.Request) fake.Step {
-		if req.ToolChoice == "none" {
-			return fake.Reply("   ") // finalization produces only whitespace
-		}
-		n++
-		return spinToolCall(n)
-	}))
-	mdl, err := p.Model("mock")
-	if err != nil {
-		t.Fatal(err)
-	}
-	fs, _ := newRepoFS(t.TempDir(), "diff --git a/x b/x\n+y\n")
-
-	if _, err := runAgent(context.Background(), mdl, fs, "sys", "task", 4); err == nil {
-		t.Error("runAgent should error when the finalization fallback also yields no output")
 	}
 }
 

entrypoint.sh

@@ -45,6 +45,10 @@
 #   GADFLY_FINDINGS_URL   optional gadfly-reports store base URL; set to POST the run +
 #                         findings for model-quality tracking (off when empty)
 #   GADFLY_FINDINGS_TOKEN optional bearer token for the gadfly-reports store
+#   GADFLY_CONSOLIDATE    cross-model consensus comment: "auto" (default; on for >=2
+#                         models), "1" force on, "0" force off (one comment per model)
+#   GADFLY_INLINE_REVIEW  when consolidating, also post a COMMENT-state PR review with
+#                         inline comments on changed lines (default on; "0" disables)
 set -uo pipefail
 
 # One model by default: the specialist suite already provides breadth, so a
@@ -64,6 +68,29 @@ die() { log "ERROR: $*"; exit 1; }
 
 API() { curl -fsS --connect-timeout 20 --max-time 30 -H "Authorization: token ${GITEA_TOKEN}" "$@"; }
 
+# upsert_comment_body MARKER BODY — create or update (by leading MARKER) a single
+# PR comment. Mirrors run.sh's per-model upsert; used for the consensus comment
+# and the per-model fallback when consolidation is on.
+upsert_comment_body() {
+  local marker="$1" body="$2" post_body existing_id="" page=1 cmts
+  post_body="$(jq -n --arg b "$body" '{body:$b}')"
+  while [ "$page" -le 10 ]; do
+    cmts="$(API "${GITEA_API}/issues/${PR}/comments?limit=50&page=${page}" || echo '[]')"
+    [ "$(echo "$cmts" | jq 'length')" = "0" ] && break
+    existing_id="$(echo "$cmts" | jq -r --arg m "$marker" \
+      '.[] | select(.body != null and (.body | startswith($m))) | .id' | head -n1)"
+    [ -n "$existing_id" ] && break
+    page=$((page+1))
+  done
+  if [ -n "$existing_id" ]; then
+    curl -sS --connect-timeout 20 --max-time 30 -X PATCH -H "Authorization: token ${GITEA_TOKEN}" \
+      -H "Content-Type: application/json" "${GITEA_API}/issues/comments/${existing_id}" -d "$post_body" >/dev/null
+  else
+    curl -sS --connect-timeout 20 --max-time 30 -X POST -H "Authorization: token ${GITEA_TOKEN}" \
+      -H "Content-Type: application/json" "${GITEA_API}/issues/${PR}/comments" -d "$post_body" >/dev/null
+  fi
+}
+
 # --- is the commenter allowed to trigger a re-review? ----------------------
 actor_allowed() {
   local actor="$1"
@@ -156,6 +183,36 @@ export GADFLY_FINDINGS_TOKEN="${GADFLY_FINDINGS_TOKEN:-}"
 MODELS="${GADFLY_MODELS:-${OLLAMA_REVIEW_MODELS:-$DEFAULT_MODELS}}"
 DEFAULT_CONC="${GADFLY_CONCURRENCY:-1}"
 
+# --- huge-PR downshift ------------------------------------------------------
+# A very large diff is what burns the model budget: every review step re-sends
+# it, multiplied across models × lenses × passes × steps (this is what nuked a
+# whole Ollama Cloud block on one giant PR). entrypoint is the only process that
+# spans the whole fleet, so the fleet-wide size decision lives here: size the PR
+# diff ONCE, and above GADFLY_HUGE_DIFF_BYTES collapse to a single cheap model +
+# a focused lens subset, fewer steps, no recheck, and a smaller embedded diff.
+# A finished shallow review beats a budget-nuking one. All knobs override; set
+# GADFLY_HUGE_DIFF_BYTES=0 to disable. Small PRs are never touched.
+HUGE_PR=0
+HUGE_DIFF_BYTES="${GADFLY_HUGE_DIFF_BYTES:-600000}"
+if [ "$HUGE_DIFF_BYTES" -gt 0 ] 2>/dev/null; then
+  PR_DIFF_BYTES="$(API "${GITEA_API}/pulls/${PR}.diff" 2>/dev/null | wc -c | tr -d '[:space:]')"
+  [ -z "$PR_DIFF_BYTES" ] && PR_DIFF_BYTES=0
+  if [ "$PR_DIFF_BYTES" -gt "$HUGE_DIFF_BYTES" ] 2>/dev/null; then
+    HUGE_PR=1
+    log "huge PR: diff ${PR_DIFF_BYTES}B > ${HUGE_DIFF_BYTES}B — downshifting the fleet (advisory)"
+    MODELS="${GADFLY_HUGE_DIFF_MODELS:-${MODELS%%,*}}" # first model only by default
+    export GADFLY_SPECIALISTS="${GADFLY_HUGE_DIFF_SPECIALISTS:-security,correctness,error-handling}"
+    export GADFLY_MAX_STEPS="${GADFLY_HUGE_DIFF_MAX_STEPS:-12}"
+    export GADFLY_RECHECK_MAX_STEPS="${GADFLY_HUGE_DIFF_RECHECK_MAX_STEPS:-8}"
+    export GADFLY_RECHECK="${GADFLY_HUGE_DIFF_RECHECK:-0}"               # skip recheck on huge PRs
+    # The Go-visible name directly (run.sh prefers GADFLY_MAX_DIFF_CHARS over its
+    # own MAX_DIFF_CHARS), so the cap is honored without relying on run.sh's alias.
+    export GADFLY_MAX_DIFF_CHARS="${GADFLY_HUGE_DIFF_MAX_DIFF_CHARS:-20000}"
+    # Surfaced on each posted comment so the shallower review is self-explaining.
+    export GADFLY_NOTICE="⚠️ Large PR (${PR_DIFF_BYTES} bytes): Gadfly downshifted to a focused, single-model review to stay within budget — coverage is intentionally shallower. Consider splitting the PR for a deeper review."
+  fi
+fi
+
 provider_of() { case "$1" in */*) echo "${1%%/*}";; *) echo "${GADFLY_PROVIDER:-ollama-cloud}";; esac; }
 
 # Per-model status file path for the live board. The model id can contain '/'
@@ -176,10 +233,11 @@ provider_cap() { # provider -> concurrency (override map "p=N,...", else default
 }
 
 review_one() {
-  local sf=""
+  local sf="" ff=""
   [ "${GADFLY_STATUS_BOARD:-1}" != "0" ] && sf="$(status_file_for "$1")"
+  [ "$CONSOLIDATE" = "1" ] && ff="$(findings_file_for "$1")"
   PROVIDER=ollama MODEL="$1" GADFLY_BIN="/usr/local/bin/gadfly" GADFLY_REPO_DIR="$REPO_DIR" \
-    GADFLY_STATUS_FILE="$sf" \
+    GADFLY_STATUS_FILE="$sf" GADFLY_FINDINGS_OUT="$ff" GADFLY_CONSOLIDATE="$CONSOLIDATE" \
     bash "${SCRIPTS_DIR}/run.sh" || log "model $1 failed (continuing)"
   # If the binary never wrote real status (run.sh skipped it: empty diff, no key,
   # binary missing), the pre-seed stays {started:0, done:false} and the board
@@ -196,6 +254,34 @@ IFS=',' read -ra _raw <<< "$MODELS" || true
 MODEL_LIST=()
 for raw in "${_raw[@]}"; do m="$(echo "$raw" | tr -d '[:space:]')"; [ -n "$m" ] && MODEL_LIST+=("$m"); done
 
+# --- cross-model consolidation decision ------------------------------------
+# With >=2 models, post ONE consensus comment (findings clustered + ranked by
+# cross-model agreement) instead of N per-model walls of prose. Each model writes
+# its findings to FINDINGS_DIR; a final pass (the binary in GADFLY_CONSOLIDATE_DIR
+# mode) renders the consensus comment. GADFLY_CONSOLIDATE: "auto" (default; on for
+# >=2 models), "1" force on, "0" force off (keep per-model comments).
+FINDINGS_DIR="${WORKDIR}/findings"
+CONSOLIDATE=0
+case "${GADFLY_CONSOLIDATE:-auto}" in
+  1) CONSOLIDATE=1 ;;
+  0) CONSOLIDATE=0 ;;
+  *) [ "${#MODEL_LIST[@]}" -ge 2 ] && CONSOLIDATE=1 ;;
+esac
+# A model spec can contain '/' and ':' (e.g. claude-code/opus, qwen3:14b), so
+# sanitize to a flat filename — but append a checksum of the raw spec so two
+# specs that sanitize the same (foo:bar vs foo/bar -> foo_bar) don't collide onto
+# one file and silently drop a model from the consensus.
+findings_file_for() {
+  local safe sum
+  safe="$(echo "$1" | tr -c '[:alnum:]._-' '_')"
+  sum="$(printf '%s' "$1" | cksum | cut -d' ' -f1)"
+  echo "${FINDINGS_DIR}/${safe}-${sum}.json"
+}
+if [ "$CONSOLIDATE" = "1" ]; then
+  rm -rf "$FINDINGS_DIR"; mkdir -p "$FINDINGS_DIR"
+  log "consolidation ON: ${#MODEL_LIST[@]} models -> one consensus comment"
+fi
+
 # Distinct providers, in first-seen order (no associative arrays — portable).
 PROVIDERS=""
 for m in "${MODEL_LIST[@]}"; do
@@ -241,6 +327,32 @@ if [ "${GADFLY_STATUS_BOARD:-1}" != "0" ]; then
   log "status board started (pid ${BOARD_PID})"
 fi
 
+# --- swarm-wide hard backstop ----------------------------------------------
+# A wall-clock ceiling across the WHOLE fleet, so a pathological PR can never
+# drain the usage block however the models behave. entrypoint is the only
+# process spanning every model, so a single "never exceed X" guard lives here.
+# On expiry it stops the review subtrees (the binary + run.sh); whatever partial
+# findings were gathered are still posted and the job never fails (advisory).
+# GADFLY_PR_BUDGET_SECS=0 (default) disables it.
+KILLER_PID=""
+rm -f "${WORKDIR}/.budget_killed" "${WORKDIR}/.disarmed" 2>/dev/null || true
+if [ "${GADFLY_PR_BUDGET_SECS:-0}" -gt 0 ] 2>/dev/null; then
+  (
+    sleep "${GADFLY_PR_BUDGET_SECS}"
+    log "PR wall-clock budget (${GADFLY_PR_BUDGET_SECS}s) reached — stopping the review fleet (advisory; partial findings still posted)"
+    : > "${WORKDIR}/.budget_killed"
+    pkill -TERM -f '/usr/local/bin/gadfly' 2>/dev/null || true
+    pkill -TERM -f "${SCRIPTS_DIR}/run.sh" 2>/dev/null || true
+    sleep 5
+    # Guard the delayed SIGKILL on the disarm marker: once the lanes finished and
+    # the watchdog was disarmed, the consolidation gadfly pass runs next, and a
+    # name-based KILL here must NOT catch it.
+    [ -f "${WORKDIR}/.disarmed" ] || pkill -KILL -f '/usr/local/bin/gadfly' 2>/dev/null || true
+  ) &
+  KILLER_PID=$!
+  log "PR budget watchdog armed (${GADFLY_PR_BUDGET_SECS}s, pid ${KILLER_PID})"
+fi
+
 log "providers: ${PROVIDERS:-none}"
 # Each provider lane runs in parallel; cap is enforced within each lane. Track
 # the lane PIDs so we wait ONLY for the review work — not the status board,
@@ -252,9 +364,63 @@ for p in $PROVIDERS; do
 done
 [ "${#LANE_PIDS[@]}" -gt 0 ] && wait "${LANE_PIDS[@]}"
 
+# Reviews finished (or the watchdog killed them): disarm the watchdog so its
+# delayed SIGKILL can't catch the consolidation pass that runs next. Drop the
+# disarm marker FIRST so even a racing watchdog that already reached its KILL line
+# skips it (the kill below also tears the watchdog subshell down during its sleep).
+if [ -n "$KILLER_PID" ]; then
+  : > "${WORKDIR}/.disarmed"
+  kill "$KILLER_PID" 2>/dev/null || true
+fi
+
+# If the backstop fired, note it on the consensus comment (per-model comments
+# were already posted during the run; a killed model surfaces as a failed lane).
+if [ -f "${WORKDIR}/.budget_killed" ]; then
+  export GADFLY_NOTICE="${GADFLY_NOTICE:+${GADFLY_NOTICE} }⏱️ This review was stopped early by the per-PR time budget (GADFLY_PR_BUDGET_SECS); findings are partial."
+fi
+
 # Reviews are done: signal the board to render the final state once and exit.
 if [ -n "$BOARD_PID" ]; then
   touch "${STATUS_DIR}/.done" 2>/dev/null || true
   wait "$BOARD_PID" 2>/dev/null || true
 fi
+
+# --- cross-model consensus comment -----------------------------------------
+# Render ONE consensus comment from the per-model findings the swarm wrote. This
+# is advisory and best-effort: if the consolidation pass produces nothing, fall
+# back to posting each model's review as its own comment (the per-model comments
+# were suppressed during the run), so a consolidation hiccup never loses output.
+if [ "$CONSOLIDATE" = "1" ]; then
+  n_files="$(ls -1 "${FINDINGS_DIR}"/*.json 2>/dev/null | wc -l | tr -d '[:space:]')"
+  log "consolidating findings from ${n_files} model(s)"
+  # Fetch the PR diff so the consolidator can also post an inline PR review,
+  # anchoring findings to changed lines (GADFLY_DIFF_FILE). GITEA_API/GITEA_TOKEN/
+  # GADFLY_PR are already in the binary's environment. Best-effort: an empty diff
+  # just means no inline review.
+  DIFF_FILE="${WORKDIR}/pr.diff"
+  API "${GITEA_API}/pulls/${PR}.diff" > "$DIFF_FILE" 2>/dev/null || true
+  CONSENSUS="$(GADFLY_CONSOLIDATE_DIR="$FINDINGS_DIR" GADFLY_DIFF_FILE="$DIFF_FILE" \
+    /usr/local/bin/gadfly 2>"${WORKDIR}/consolidate.err" || true)"
+  if [ -n "$CONSENSUS" ]; then
+    NOTICE_BLOCK=""
+    [ -n "${GADFLY_NOTICE:-}" ] && NOTICE_BLOCK="> ${GADFLY_NOTICE}"$'\n\n'
+    BODY="$(printf '%s%s\n\n<sub>Automated adversarial review by Gadfly — consensus across the model swarm. Advisory only — does not block merge.</sub>' "$NOTICE_BLOCK" "$CONSENSUS")"
+    upsert_comment_body "<!-- gadfly-consensus -->" "$BODY"
+    log "consensus comment posted"
+  else
+    log "consolidation produced no output; falling back to per-model comments"
+    log "$(tail -c 500 "${WORKDIR}/consolidate.err" 2>/dev/null)"
+    for f in "${FINDINGS_DIR}"/*.json; do
+      [ -f "$f" ] || continue
+      m="$(jq -r '.model // ""' "$f" 2>/dev/null)"
+      [ -z "$m" ] && continue
+      prov="$(jq -r '.provider // ""' "$f" 2>/dev/null)"
+      md="$(jq -r '.markdown // ""' "$f" 2>/dev/null)"
+      marker="<!-- gadfly-review:ollama:${m} -->"
+      body="$(printf '%s\n### 🪰 Gadfly review — `%s` (%s)\n\n%s\n\n<sub>Automated adversarial review by Gadfly. Advisory only — does not block merge.</sub>' \
+        "$marker" "$m" "$prov" "$md")"
+      upsert_comment_body "$marker" "$body"
+    done
+  fi
+fi
 log "done"

examples/reusable.yml

@@ -14,10 +14,11 @@
 # Forward ONLY the secrets the reviewer uses (least privilege) — see the
 # `secrets:` block below. GITEA_TOKEN is automatic. `secrets: inherit` also works
 # but hands the reusable EVERY secret in your repo (registry/deploy/db creds the
-# review never touches), so prefer the explicit form. Pin @<ref>: use the @v1
+# review never touches), so prefer the explicit form. Pin to an immutable
-# release tag (a curated pointer moved on deliberate releases) for auto-updating
+# @<sha>: long-lived act_runners CACHE the reusable by ref, so a moved tag (@v1)
-# stability, or a full @<sha> for an immutable pin. Avoid @main — it moves on
+# or @main is often not re-fetched and silently runs a stale copy. Bump the @<sha>
-# every push and would change what runs with your forwarded secrets.
+# to adopt a structural change; routine swarm tuning rides owner variables (see
+# the gadfly README "Central config via variables") with no re-pin needed.
 #
 # For custom named endpoints (GADFLY_ENDPOINT_<NAME>) or a provider the reusable
 # doesn't map, use the full stub in adversarial-review.yml instead.
@@ -49,8 +50,8 @@ jobs:
     if: >-
       github.event_name != 'issue_comment'
       || (github.event.issue.pull_request && github.actor == 'your-username')
-    # @v1 = curated release tag (auto-updates on releases); swap for a full @<sha>
+    # Pin to an immutable @<sha> (runners cache the ref, so @v1/@main can run
-    # if you want an immutable pin. Don't use @main (moves on every push).
+    # stale). Bump it for structural changes; tune the swarm via owner variables.
     uses: steve/gadfly/.gitea/workflows/review-reusable.yml@v1
     # Forward ONLY what the reviewer needs. Add provider keys you use
     # (ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY, GADFLY_API_KEY) and/or

go.mod

@@ -3,6 +3,7 @@ module gitea.stevedudenhoeffer.com/steve/gadfly
 go 1.26.2
 
 require (
+	gitea.stevedudenhoeffer.com/steve/executus v0.1.4
 	gitea.stevedudenhoeffer.com/steve/majordomo v0.0.0-20260627225659-aa25b2c33462
 	gopkg.in/yaml.v3 v3.0.1
 )
@@ -17,6 +18,7 @@ require (
 	github.com/go-logr/stdr v1.2.2 // indirect
 	github.com/google/go-cmp v0.7.0 // indirect
 	github.com/google/s2a-go v0.1.9 // indirect
+	github.com/google/uuid v1.6.0 // indirect
 	github.com/googleapis/enterprise-certificate-proxy v0.3.17 // indirect
 	github.com/googleapis/gax-go/v2 v2.22.0 // indirect
 	github.com/gorilla/websocket v1.5.3 // indirect

go.sum

@@ -4,6 +4,8 @@ cloud.google.com/go/auth v0.20.0 h1:kXTssoVb4azsVDoUiF8KvxAqrsQcQtB53DcSgta74CA=
 cloud.google.com/go/auth v0.20.0/go.mod h1:942/yi/itH1SsmpyrbnTMDgGfdy2BUqIKyd0cyYLc5Q=
 cloud.google.com/go/compute/metadata v0.9.0 h1:pDUj4QMoPejqq20dK0Pg2N4yG9zIkYGdBtwLoEkH9Zs=
 cloud.google.com/go/compute/metadata v0.9.0/go.mod h1:E0bWwX5wTnLPedCKqk3pJmVgCBSM6qQI1yTBdEb3C10=
+gitea.stevedudenhoeffer.com/steve/executus v0.1.4 h1:4F99uCV3OVaE9ITFp0FjPiYxLUQO+WpE+wU2HCnpXNM=
+gitea.stevedudenhoeffer.com/steve/executus v0.1.4/go.mod h1:WQP/lH+meU06OSNF0TQO/wQLcJCrMwpi0EMj5vSpVtk=
 gitea.stevedudenhoeffer.com/steve/majordomo v0.0.0-20260627225659-aa25b2c33462 h1:1crjE1YkWHLZ91tUDOxN/Y5cuOnJ56e0U9UADoFfEPY=
 gitea.stevedudenhoeffer.com/steve/majordomo v0.0.0-20260627225659-aa25b2c33462/go.mod h1:UZLveG17SmENt4sne2RSLIbioix30RZbRIQUzBAnOyY=
 github.com/cespare/xxhash/v2 v2.3.0 h1:UL815xU9SqsFlibzuggzjXhog7bL6oX9BbNZnL2UFvs=

scripts/run.sh

@@ -50,6 +50,13 @@ MAX_DIFF_CHARS="${MAX_DIFF_CHARS:-60000}"
 MARKER="<!-- gadfly-review:${PROVIDER}:${MODEL} -->"
 say() { echo "[gadfly-review:${PROVIDER}:${MODEL}] $*" >&2; }
 
+# When the swarm is consolidating (GADFLY_CONSOLIDATE=1, set by entrypoint.sh for
+# a multi-model run), this model does NOT post its own comment — it writes its
+# findings to GADFLY_FINDINGS_OUT and a single cross-model consensus comment is
+# posted after the whole swarm finishes. Live progress still shows on the status
+# board. Default 0 (post a per-model comment, the standalone behavior).
+CONSOLIDATE="${GADFLY_CONSOLIDATE:-0}"
+
 # Display the model's ACTUAL backend: the provider segment of the spec
 # ("m1pro/qwen3.6:35b-mlx" -> "m1pro"); a bare id uses GADFLY_PROVIDER (default
 # ollama-cloud). This is what the comment header shows, not the run.sh lane.
@@ -126,7 +133,9 @@ USR="$(printf 'PR #%s: %s\n\nDescription:\n%s\n\nUnified diff to review:\n```dif
 # --- announce start (placeholder comment) -----------------------------------
 START_TS="$(date +%s)"
 say "starting review with ${MODEL}"
-upsert_comment "$(printf '%s\n### 🪰 Gadfly review — `%s` (%s)\n\n⏳ Reviewing… this comment will update with findings and run time.' \
+# Skip the per-model placeholder when consolidating (the consensus comment is
+# posted later; live progress is on the status board).
+[ "$CONSOLIDATE" = "1" ] || upsert_comment "$(printf '%s\n### 🪰 Gadfly review — `%s` (%s)\n\n⏳ Reviewing… this comment will update with findings and run time.' \
   "$MARKER" "$MODEL" "$MODEL_PROVIDER")"
 
 # --- call the model ---------------------------------------------------------
@@ -168,8 +177,9 @@ case "$PROVIDER" in
           GADFLY_SYSTEM_FILE="${SCRIPT_DIR}/system-prompt.txt" \
           GADFLY_TITLE="$TITLE" \
           GADFLY_BODY="$BODY" \
-          GADFLY_MAX_DIFF_CHARS="$MAX_DIFF_CHARS" \
+          GADFLY_MAX_DIFF_CHARS="${GADFLY_MAX_DIFF_CHARS:-$MAX_DIFF_CHARS}" \
           GADFLY_STATUS_FILE="${GADFLY_STATUS_FILE:-}" \
+          GADFLY_FINDINGS_OUT="${GADFLY_FINDINGS_OUT:-}" \
           "$BIN" 2>"$ERR_FILE"
         )"
         rc=$?
@@ -204,7 +214,28 @@ esac
 # --- assemble + post final comment (with run time) --------------------------
 ELAPSED="$(( $(date +%s) - START_TS ))"
 DUR="$(fmt_duration "$ELAPSED")"
-COMMENT="$(printf '%s\n### 🪰 Gadfly review — `%s` (%s)\n\n%s\n\n<sub>Automated adversarial review by Gadfly. Advisory only — does not block merge. · ⏱️ reviewed in %s</sub>' \
+
-  "$MARKER" "$MODEL" "$MODEL_PROVIDER" "$REVIEW" "$DUR")"
+# Consolidating: the binary writes its findings file on success. If it failed or
+# was skipped (no file, or an empty one), write a stub so this model still shows
+# up in the consensus (as failed) and an all-models-fail run still posts a
+# comment — never silently drop a model or the whole review.
+if [ "$CONSOLIDATE" = "1" ] && [ -n "${GADFLY_FINDINGS_OUT:-}" ] && [ ! -s "${GADFLY_FINDINGS_OUT}" ]; then
+  jq -n --arg model "$MODEL" --arg provider "$MODEL_PROVIDER" --arg md "$REVIEW" \
+    '{model:$model, provider:$provider, verdict:"reviewer failed", errored:true, markdown:$md, findings:[]}' \
+    > "${GADFLY_FINDINGS_OUT}" 2>/dev/null || true
+fi
+# When consolidating, the binary has written this model's findings to
+# GADFLY_FINDINGS_OUT; the consensus comment is posted by entrypoint.sh after the
+# whole swarm finishes, so this model posts no comment of its own.
+if [ "$CONSOLIDATE" = "1" ]; then
+  say "done in ${DUR} (consolidated; no per-model comment)"
+else
+  # An optional one-line notice (e.g. entrypoint's huge-PR downshift advisory),
+  # shown under the header so a shallower review is self-explaining.
+  NOTICE_BLOCK=""
+  [ -n "${GADFLY_NOTICE:-}" ] && NOTICE_BLOCK="> ${GADFLY_NOTICE}"$'\n\n'
+  COMMENT="$(printf '%s\n### 🪰 Gadfly review — `%s` (%s)\n\n%s%s\n\n<sub>Automated adversarial review by Gadfly. Advisory only — does not block merge. · ⏱️ reviewed in %s</sub>' \
+    "$MARKER" "$MODEL" "$MODEL_PROVIDER" "$NOTICE_BLOCK" "$REVIEW" "$DUR")"
   upsert_comment "$COMMENT"
   say "done in ${DUR}"
+fi

scripts/system-prompt.txt

@@ -16,7 +16,9 @@ state. USE THEM to verify before you report. Do not review the diff in isolation
 - list_dir([path]) — list a directory.
 - grep(pattern[, path, max_results]) — RE2 regex search across the repo.
 - find_files(name[, max_results]) — locate a file by path substring.
-- get_diff() — the full unified diff (the task message may truncate it).
+- get_diff([path, start_line, limit]) — the unified diff as a paginated, numbered window;
+  pass `path` to fetch just one changed file's hunks (do this on a big PR instead of pulling
+  the whole diff at once).
 
 Mandatory verification discipline — this is the whole point of giving you tools:
 - Before claiming a missing/duplicate import, an undefined symbol, a wrong signature,
@@ -43,3 +45,21 @@ Output rules:
 - Only report issues you are reasonably confident are real after checking. If the diff
   is clean, say so plainly rather than inventing nits.
 - When you are done investigating, STOP calling tools and reply with the final review.
+
+Machine-readable findings — AFTER the prose review, append ONE fenced code block,
+tagged `gadfly-findings`, holding a JSON array of the SAME findings you described above
+(this block is consumed by tooling and hidden from the rendered comment):
+
+```gadfly-findings
+[
+  {"file": "path/to/file.go", "line": 123, "severity": "high", "confidence": "high", "title": "one-line summary of the issue"}
+]
+```
+
+- One object per real finding, in the same order as your prose. `file`/`line` must be a
+  concrete location you verified (the line the issue is at). `severity` is one of
+  `critical`, `high`, `medium`, `small`, `trivial`. `confidence` is your post-verification
+  confidence the issue is real: one of `high`, `medium`, `low`.
+- Include ONLY genuine problems — never verification notes ("confirmed X is safe at f:line"),
+  and never an "Outside my lens:" aside. If your lens is clean, emit an empty array `[]`.
+- This block is in ADDITION to the prose; do not drop the human-readable findings.