feat: re-platform agentic review onto executus + large-PR cost controls
Build & push image / build-and-push (pull_request) Successful in 11s
Adversarial Review (Gadfly) / review (pull_request) Successful in 25m56s

Fixes the large-PR token burn: a ~250K-token diff was re-sent every agent
step across models × lenses × passes, draining a metered usage block in
minutes. Small PRs are untouched (every mitigation is size-gated / no-op
under threshold).

- Re-platform the in-process review path onto executus run.Executor: context
  compaction (executus/compact, threshold from the model's real context window
  via executus/model), run-bounding, a per-PR budget gate (Ports.Budget), and
  the wrap-up nudge re-expressed as a run.Critic. Lens fan-out now uses
  executus/fanout. gadfly keeps its own model.go, so GADFLY_ENDPOINT_<NAME>
  aliases and the claude-code engine are unaffected. No majordomo bump; the
  binary stays static (executus core is majordomo+stdlib only).
- Paginate get_diff (per-file `path` + start_line/limit) instead of dumping the
  whole diff; trim the recheck diff embed (60k -> 20k chars).
- entrypoint.sh: downshift the fleet above GADFLY_HUGE_DIFF_BYTES (one cheap
  model, fewer lenses/steps, no recheck) + a swarm-wide GADFLY_PR_BUDGET_SECS
  wall-clock backstop (adds procps for pkill). All advisory; CI never fails.
- README + CLAUDE.md + tests updated.

Note: run.Result exposes no transcript, so the old transcript-based forced-
finalization fallback is dropped; the wrap-up critic nudge is the remaining
"always emit something" mechanism.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-06-30 10:43:34 -04:00
parent 5007597cf9
commit b860119332
19 changed files with 935 additions and 224 deletions
+4 -1
View File
@@ -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