Fourteen findings, and seven of them from all four models are the same one:
endpointProvider was missing the no-cross-vendor-fallback guard I had just
added to resolveModel. I fixed a credential leak on one path and left its
sibling leaking, in the commit whose own message argued those two paths must
move together. That is the third time in this PR.
So it is no longer a rule written twice. openAICompatOptions owns it and both
paths call it; builtinCompatProviders names the vendors that must never inherit
OPENAI_API_KEY, replacing a `provider == "kimi" || provider == "qwen"` literal
that was a fourth uncounted copy of the list.
The test drives a real request at a local server and demands two things: that
no request arrives carrying the OpenAI key, AND that the call fails closed
naming the variable to set — the second half because my first draft pointed the
provider at vendor.example, so the server saw nothing and the assertion held
for a reason unrelated to the fix. Break-checked: removing the guard puts
"Bearer sk-openai-must-not-travel" on the wire to the other vendor.
The scrub check failed open. As a bare condition, a grep ERROR (exit >= 2)
reads as "not found" and skips the guard — a credential check that passes
precisely when it cannot see the filesystem it is searching. It now
distinguishes 0/1/>=2 and refuses to continue on error.
A bare "claude-code" spec has no "/", so the provider fell back to ollama-cloud
and the pre-flight would skip a reviewer that authenticates with
CLAUDE_CODE_OAUTH_TOKEN and needs no Ollama key. Engine specs are now exempt.
preflight.sh's provider list duplicated its own case arms; both now read one
table. And its comment claimed the Go cross-check fails if either list misses
an entry from the other, when only one direction is checked — the reverse is
not even desirable, since ollama-cloud and anthropic belong in that table and
not in the Go one. The comment now says what is enforced.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Round 6, and one finding exposed something no reviewer mentioned: the majordomo
bump this whole PR depends on was never made. Every test here builds the openai
client directly, so all of them passed against a majordomo release that had
never heard of qwen — a plain "qwen/<model>" in GADFLY_MODELS, the primary way
anyone will use this, would not have resolved at all. A compile error caught it,
which is luck. TestBuiltinCompatProvidersResolveViaRegistry now exercises that
path; the build is what guards the dep itself, since the old release cannot
compile the code below.
On the endpoint-override path, kimi and qwen fell through to openai.New's
OPENAI_API_KEY default whenever GADFLY_API_KEY was unset — sending an OpenAI
key to Moonshot or Alibaba. That is a credential handed to the wrong vendor,
and it is the exact failure majordomo's built-ins are written to prevent; I
reintroduced it one layer up. Both now pass the key unconditionally, so an
absent key is a 401 naming GADFLY_API_KEY rather than a foreign credential on
the wire.
The test job scrubbed the registry credential and left the checkout token in
.git/config, readable by the `go test` it then runs — fixing one credential
while its neighbour sat in the open. persist-credentials: false; nothing in
that job talks to git after checkout.
The cross-language wiring test now QUERIES preflight.sh via a new
gadfly_preflight_providers function instead of regexing its case statement.
Parsing made that file's formatting a contract no linter enforces, where a
harmless reformat breaks a test in another language. Two models flagged it.
Also: grep for the scrub check takes -e, so a password starting with a hyphen
is not read as options; and key_hint stopped repeating key_env in four of five
arms.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Round 5, and the best findings are again about the fix from round 4.
The scrub only ran on success. `set -e` aborts the step when `go mod download`
fails, so the cleanup line after it never executed — leaving a push-capable
credential on a long-lived self-hosted runner for whatever job landed there
next. It is now a `trap ... EXIT`, verified against a simulated failure.
It also scrubbed the wrong file in principle: `git config --global` writes to
GIT_CONFIG_GLOBAL, else $XDG_CONFIG_HOME/git/config when that exists, else
~/.gitconfig — so deleting ~/.gitconfig can scrub a path the credential was
never in. The step now names GIT_CONFIG_GLOBAL itself, leaving exactly one file
to remove.
And the verification failed open in the case that matters most: `grep -F ""`
matches every file, so a run WITHOUT the secret — a fork PR, the threat model —
failed the check with a message accusing it of leaking a credential it never
had. Guarded on a non-empty secret.
Credentials move to an Authorization header instead of being embedded in the
URL, so a password containing @ : / or # can no longer break URL parsing in a
way that reads as a bad password.
Two list-drift holes closed with one test that reads across languages:
TestOpenAICompatProvidersAreFullyWired asserts every openAICompatProviders
entry is both advertised in endpointProviderNames and has a credential arm in
scripts/preflight.sh. Adding a compat provider touches three places in two
languages and nothing connected them. Break-checked in both directions.
Finally, a whitespace-only GADFLY_BASE_URL disagreed across the boundary: Go
TrimSpaces it and takes the registry path, bash called it "set" and skipped the
pre-flight, so the missing key arrived as a bare 401 with no notice. Both now
agree on what unset means.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Both Claude reviewers caught this independently, and they are right. The test
job I added wrote a PUSH-CAPABLE REGISTRY_PASSWORD into a plaintext
~/.gitconfig and then ran `go build`/`go vet`/`go test` — repository code — on
pull_request events. This repo is public, so a fork PR could ship a test whose
only job is to print that file. The image build had already answered this
question correctly: its credentials are BuildKit secrets scoped to the
module-download RUN and are never present while code executes. I bolted on a
job that skipped the boundary its neighbour maintains.
Dependencies are now fetched in their own step which deletes ~/.gitconfig
before anything else runs, and asserts the scrub — against the whole home
directory, not against the file it just removed, because the credential can
also land in ~/.netrc or ~/.config/go/env. Verified the assertion is not
vacuous: planting the secret in ~/.netrc trips it. Later steps run with
GOPROXY=off, so any attempt to reach the network fails loudly rather than
quietly hunting for the credential that is now gone.
Also from round 4: TestEndpointProviderNamesAreAllAccepted pinned only
endpointProvider, while the constant is the error text for BOTH resolution
paths — it now asserts each advertised name resolves either way (break-checked
by dropping the gemini alias from resolveModel alone). preflight.sh documents
that ollama-cloud is checked on OLLAMA_API_KEY but hinted as
OLLAMA_CLOUD_API_KEY because run.sh copies one to the other first, an ordering
dependency that was invisible from the file.
And the comments that narrated this PR's own edit history ("the first version
of this change...") are rewritten as invariants. That history stops being true
the moment this merges, and the repo's doc policy says as much.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Twelve findings, all real, and the two that matter are about the pre-flight I
added rather than about qwen.
The credential check had a false pass in the OTHER direction from round 2's: on
the GADFLY_BASE_URL override path, resolveModel builds the client with
GADFLY_API_KEY and never reads QWEN_API_KEY/KIMI_API_KEY, so treating the
provider's own key as sufficient there let a doomed run proceed. Having now
been wrong about these rules in both directions, the check no longer tries to
model both paths: it covers the REGISTRY path, whose rules it can state
exactly, and says nothing about the override path — which is hand-configured by
definition, while the registry path is the one you hit by adding a model id to
a var and forgetting the secret.
The logic moves to scripts/preflight.sh, sourced by both run.sh and the test.
The previous answer to "this test duplicates production logic" was a regex
drift-guard, and that guard compared only the provider table — not the decision
logic, which is precisely the half that carried the bug. A duplicate you guard
is still a duplicate; this deletes it, and the test now runs under `set -u`
like production does.
Also: the test that pins the shared provider slice held its own copy of the
list (now ranges the slice); endpointProviderNames had nothing tying it to the
switches it describes, which is how it shipped without "gemini" (a new test
asserts every advertised name resolves); two godoc lists had drifted; and the
"sanity" line that asserted nothing is gone.
And the repo had NO test job — `go test` and the pre-flight table both existed
and neither was ever executed by CI, which reads as coverage while providing
none. Added one (build/vet/gofmt/test/pre-flight), running alongside the image
build rather than gating it, so red is loud without standing between a push and
a rebuild.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Eight findings, all real, and the sharpest ones are about this PR's own fixes.
GADFLY_API_KEY was treated as a universal substitute in the pre-flight. It is
not: resolveModel reads it only AFTER the `baseURL == ""` early return, so on
the registry path — the documented primary path — a qwen/kimi built-in reads
its own variable and GADFLY_API_KEY is never consulted. A mis-set
GADFLY_API_KEY therefore passed pre-flight and 401'd five times anyway, which
is precisely the failure this check exists to prevent. It now only substitutes
when GADFLY_BASE_URL is also set.
`openai-compatible` was missing from the pre-flight table while both switches
accept it as an OPENAI_API_KEY alias, so that one spelling still fell through
to the cryptic five-failure mode.
endpointProviderNames — the constant I introduced *to stop* the two error
messages drifting — omitted the `gemini` alias both switches accept. It now
lists every accepted spelling.
And the case list itself was still duplicated across both switches plus the
test that pins them: three copies of the thing whose duplication started this.
Both switches now call isOpenAICompatProvider over one shared slice, and
endpointProvider's doc comment points at endpointProviderNames instead of
carrying a fourth hand-written copy.
scripts/preflight_test.sh moves into the repo (20 cases, up from 17, covering
openai-compatible and both GADFLY_API_KEY directions). It carries a drift guard
that diffs its copy of the provider table against run.sh's and aborts if they
differ — break-checked by deleting an arm from run.sh, which fails it loudly.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The pre-flight comment was the worst of them, and three models agreed. It said
providers absent from the table "need no key or carry it in their endpoint/DSN"
— false for google, which needs a key and is absent for an entirely different
reason: it accepts GOOGLE_API_KEY *or* GEMINI_API_KEY, so a single-variable arm
would silently skip a correctly-configured reviewer. That reasoning was in the
PR description and not in the code, so the comment invited exactly the wrong
edit. It now states both exclusion reasons and names google's.
Forwarded KIMI_API_KEY alongside QWEN_API_KEY in the dogfooding stub. This PR
argues that sibling call sites must move together, and I declared both secrets
in the reusable workflow and forwarded one — a config that looks complete and
401s on the model you didn't wire.
The two endpoint-provider error messages listed the same accepted set in
different order and spelling. Both functions accept an identical set, so they
now share one endpointProviderNames constant and cannot disagree.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
majordomo now ships qwen and kimi as built-ins that ARE the openai client at
their own base URL, so "qwen/qwen3.8-max" works as a GADFLY_MODELS entry once
the key reaches the container. This wires up the parts that key has to pass
through.
Two provider switches had to learn the names, not one. resolveModel's
GADFLY_BASE_URL override was the obvious one; endpointProvider's
GADFLY_ENDPOINT_* parser is its sibling, and I fixed the first and missed the
second on the first pass — a config that resolves one way and errors the other
for no reason a user could guess. TestOpenAICompatProvidersResolveOnBothPaths
now asserts both from one table so the pair fails together; break-checked in
both directions.
QWEN_API_KEY (and KIMI_API_KEY) are declared as workflow_call secrets and
forwarded to the container, with gadfly's own stub forwarding QWEN_API_KEY so a
qwen entry can join the default swarm by editing GADFLY_DEFAULT_MODELS alone —
no workflow edit, no re-release.
The run.sh credential pre-flight is now a provider→variable table instead of an
ollama-cloud special case. Without it a forgotten key surfaces as five
identical per-lens agent failures naming no variable, and the operator reads a
stack trace to find out which secret they missed. Google stays out of the table
on purpose: it accepts either GOOGLE_API_KEY or GEMINI_API_KEY, and a one-var
entry would wrongly skip a correctly-configured run. Verified across 17
provider x key-state combinations, including that a wrong-provider key never
satisfies qwen (majordomo refuses cross-provider fallback) and that unkeyed
providers are never blocked.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Gadfly's own review of #27 surfaced a real robustness cluster (5 models,
error-handling) plus stale comments I missed.
Robustness — the flock permit pool could hang forever:
- tryAcquire swallowed os.OpenFile errors and treated every flock error as
"busy", so a broken/missing pool dir (or a filesystem without flock) would
spin-poll indefinitely; the fanout context is uncancellable and the per-lens
timeout only starts AFTER acquire returns. Can't trigger in the deploy
(entrypoint mkdir -p's the dir) but fixed defensively.
- tryAcquire now returns a structural error, distinguished from a healthy-full
pool (EWOULDBLOCK = busy → keep polling). acquire FAILS OPEN on a structural
error: logs once and runs the lens unthrottled rather than hanging the review.
- acquire uses time.NewTimer + Stop() (no per-poll timer leak on cancellation).
- activeLensSem warns on stderr when GADFLY_LENS_SEM_DIR is set but the size is
invalid (was a silent degrade to unthrottled).
- New test: a broken pool dir fails open promptly.
Doc drift (stale references to the removed model cap):
- main.go defaultLensConcurrency + runSpecialists doc, entrypoint.sh status
pre-seed + lane-launch comments, and the pre-existing lens_concurrency_test.go
header all updated to the provider-wide-budget wording.
Accepted (graded real, not changed): all-models-start-at-once startup burst
(intended tradeoff) and index-0 sweep bias (cosmetic). One false positive
(one model using the whole budget is the intended lone-model behavior).
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Concurrency was two multiplicative gates in two processes: entrypoint.sh
capped MODELS-at-once per provider (GADFLY_PROVIDER_CONCURRENCY) while each
model's binary separately capped its own lenses (GADFLY_LENS_CONCURRENCY).
A model therefore held its whole model-slot until its LAST lens finished,
stalling the next model even with idle lens capacity.
Collapse to one throttle: a provider-wide lens budget shared across all of
that provider's models. entrypoint now runs every model in a lane at once and
seeds a single cross-process permit pool per lane (a dir of N flock files,
sized by GADFLY_PROVIDER_LENS_CONCURRENCY -> GADFLY_LENS_CONCURRENCY). Each
lens pass (review+recheck) acquires a permit before it runs and releases it
after, so a model winding down immediately yields its freed permits to
another model's queued lenses. flock auto-releases on process death, so a
killed/crashed model can't leak budget.
- cmd/gadfly/lenssem.go: the flock permit pool (+ lenssem_test.go).
- main.go: runSpecialists holds a shared permit per lens; fanout sized to the
budget so a lone model can use all of it. Falls back to the in-process limit
when no pool is set (local runs, tests).
- entrypoint.sh: drop provider_cap/DEFAULT_CONC; run_lane runs all models and
seeds the per-lane pool.
- GADFLY_PROVIDER_CONCURRENCY / GADFLY_CONCURRENCY are now ignored; the
reusable workflow marks provider_concurrency deprecated and stops forwarding
it. Docs (README, CLAUDE.md, examples) updated per the maintenance rule.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Gadfly's own swarm reviewed PR #26 and reached consensus (3/3 models) on a real
bug, plus flagged security/maintainability items. Fixes:
- Pass-through auth (BLOCKING, 3/3 agreement): openCodeEnv() stripped every
provider key except OLLAMA_API_KEY, so the documented opencode/<provider>/<model>
escape hatch (e.g. opencode/anthropic/...) had no way to authenticate — the
reusable workflow forwards ANTHROPIC_API_KEY/OPENAI_API_KEY into the container
and the allowlist discarded them. Now forward ANTHROPIC_*/OPENAI_*/GOOGLE_*/
GEMINI_* so OpenCode's built-in providers can authenticate, while still
withholding gadfly's own secrets (Gitea/findings tokens, claude-code OAuth).
- Read-only hardening (security lens): the generated config denied only edit/bash.
Using OpenCode's documented permission schema, also deny webfetch/websearch/
external_directory — the network + out-of-sandbox tools — closing the
exfiltration surface a prompt-injected review could otherwise reach. Permission
is now a map so the deny set is extensible.
- Dedup (maintainability lens, 3/3): extract shared filterEnv() and
killGroupOnCancel() helpers in engine.go, used by both shell-out engines'
runPass/env builders instead of the copy-pasted blocks.
- Cosmetic: split the const block so defaultOpenCodeBaseURL's doc comment no
longer visually misattaches to the agent-name const.
README updated: the read-only note and the reduced-env note now reflect the
broader deny set and the forwarded provider keys.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Add a third review harness alongside the in-process majordomo loop and the
claude-code CLI shell-out: the OpenCode CLI (opencode.ai) driving an ollama-cloud
model, selected by an "opencode/<model>" spec. The goal is to benchmark gadfly's
boutique executus harness against a freely-available agentic harness on the SAME
model (e.g. "ollama-cloud/glm-5.2" vs "opencode/glm-5.2").
OpenCode has no --append-system-prompt flag, so the lens system prompt and the
read-only discipline are delivered through a generated config injected via
OPENCODE_CONFIG_CONTENT: a "gadfly" agent whose prompt is the system prompt with
edit/bash denied at both the global and agent level, plus a "gadfly" ollama-cloud
provider. That env var is the highest-precedence config source in the container,
so a reviewed repo's own opencode.json can't re-enable edits on the reviewer.
Spec forms: "opencode/<model>" (wrapped in the generated provider), the
"open-code/" alias, "opencode/<provider>/<model>" pass-through to OpenCode's own
registry, and bare "opencode". Model ids are taken verbatim so colon-bearing
ollama ids (qwen3-coder:480b-cloud) survive. Auth reuses OLLAMA_CLOUD_API_KEY
(mapped to OLLAMA_API_KEY, referenced as {env:OLLAMA_API_KEY} in config, never a
literal secret). Knobs mirror GADFLY_CLAUDE_*: GADFLY_OPENCODE_BIN/MODEL/BASE_URL/
EXTRA_ARGS. openCodeEnv() forwards OLLAMA_API_KEY (the inverse of claudeEnv) but
still withholds the Gitea/findings/Anthropic secrets.
main.go engine selection is now a switch (claude-code / opencode / majordomo), and
the auto-select path uses a type-check instead of a boolean so a shell-out engine
can never hit the *majordomoEngine assertion. auto-select and delegate_investigation
stay majordomo-only and are skipped for opencode (the CLI does its own legwork).
Dockerfile bundles opencode-ai (npm auto-selects its musl build on alpine) with a
best-effort version check + provider pre-warm that never fails the shared image
build. README/examples/CLAUDE.md/scripts updated per the maintenance rules.
Tests: new opencode_test.go mirrors engine_test.go (spec/model/args/config/env-
filter + stub-CLI runtime tests). Verified end-to-end with a fake opencode CLI:
correct argv, injected config, and consolidated markdown output.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
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 <[email protected]>
Co-committed-by: Steve Dudenhoeffer <[email protected]>
Bump majordomo to the latest build and accept every llama-swap spelling
(llama-swap/llama-swaps + un-hyphenated llamaswap/llamaswaps) in gadfly's
endpoint switches; the LLM_* llama-swap(s):// DSN path already worked via
majordomo.Parse. README + error messages + endpointProvider alias tests.
Swarm review: 8/9 clean; qwen3-coder's "Blocking" was a false positive
(claimed llamaswap was untested — it has dedicated test cases). Folded in
its one fair nit (README now lists the un-hyphenated aliases).
gofmt clean, go vet quiet, go test -race green.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Co-authored-by: Steve Dudenhoeffer <[email protected]>
Co-committed-by: Steve Dudenhoeffer <[email protected]>
Adds claude-code/opus to gadfly's dogfood swarm (both sonnet and opus run
end-to-end), bumps the image pin to :sha-80d8f53 so the clean-lens
telemetry fix is live, and adds engine support for a
"claude-code/<model>:max" extended-thinking spec (MAX_THINKING_TOKENS,
best-effort). Validated: only 13 findings on this clean PR vs 43 on the
comparable #4 — the telemetry fix works.
Folded in the swarm's two real findings: a runPass env-injection test and
keeping MAX_THINKING_TOKENS in claudeEnv. Follow-up enables
claude-code/opus:max once this image builds.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Co-authored-by: Steve Dudenhoeffer <[email protected]>
Co-committed-by: Steve Dudenhoeffer <[email protected]>
emit() now skips findings extraction for a "No material issues found"
lens (its path:line refs are verification notes, not problems), fixing
the FP inflation that penalized thorough clean-pass reviewers. Also trims
the dogfood swarm to the strong reviewers: drops m5/qwen3.6 (last local
lane), gemma4, gpt-oss:120b, and kimi-k2.7-code — leaving 6 cloud +
claude-code/sonnet.
Fittingly, PR #4's own 11-model review produced 43 findings that were ALL
clean-verification bullets (zero real) — a live demonstration of the bug
this fixes. gofmt clean, go vet quiet, go test -race green.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Co-authored-by: Steve Dudenhoeffer <[email protected]>
Co-committed-by: Steve Dudenhoeffer <[email protected]>
Phase 2: bump majordomo to latest and wire its new llamaswap provider
into gadfly's endpoint switches; add claude-code/sonnet to gadfly's own
dogfood swarm (pin :sha-86f12c1, map CLAUDE_CODE_OAUTH_TOKEN) so the
Phase-1 engine runs as a live competitor; document the Ollama-through-CC
ANTHROPIC_BASE_URL proxy path as example-only.
The 11-model swarm (incl. claude-code/sonnet) reviewed it; 52 findings
graded via the MCP. Folded in the two real ones: a llamaswap
endpointProvider test (caught by claude-code/sonnet, citing CLAUDE.md)
and adding "openai-compatible" to the provider error messages (gpt-oss).
gofmt clean, go vet quiet, go build + go test -race green.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Co-authored-by: Steve Dudenhoeffer <[email protected]>
Co-committed-by: Steve Dudenhoeffer <[email protected]>
Phase 1: a second review engine alongside the majordomo agent loop. For
each lens, shell out to the Claude Code CLI (`claude -p --output-format
json`) inside the checked-out repo so it verifies findings with its own
read tools, then reuse gadfly's verdict-parse + recheck + consolidate +
emit pipeline. Select via GADFLY_MODELS `claude-code`/`claude-code/<model>`;
auth via CLAUDE_CODE_OAUTH_TOKEN (no --bare) else ANTHROPIC_API_KEY;
read-only by default; GADFLY_CLAUDE_* knobs. Dockerfile bundles Node +
@anthropic-ai/claude-code. Also bumped the dogfood pin to the status-board
image (PR #2 was the first dogfood with the live board + full fleet).
Folded in the swarm's own review findings: minimal subprocess env (no
GITEA_TOKEN leak to the CLI), runPass robustness (ctx/empty-result/runErr),
process-group cleanup on timeout, rune-safe error truncation, and
engine-neutral prompts (also de-mort-ified the recheck prompt). 66 findings
graded via the gadfly MCP.
gofmt clean, go vet quiet, go build + go test -race green.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Co-authored-by: Steve Dudenhoeffer <[email protected]>
Co-committed-by: Steve Dudenhoeffer <[email protected]>
Phase 3: one consolidated, live-updating PR comment aggregating every
model's per-lens progress (queued -> running -> finished + verdict), so
the swarm's progress is visible at a glance and a watcher can tell when
it's done. Opt-in statusWriter in the binary (atomic writes) + a
background status-board.sh renderer wired through entrypoint.sh; default
on, GADFLY_STATUS_BOARD=0 to disable.
Also restores gadfly's dogfood swarm to the full cloud fleet (9 cloud +
M5; M1 dropped as too slow) matching mort, and folds in the 3 real bugs
the swarm found on its own PR (skip-binary stuck-waiting, panic-stuck
lens, busy-loop on bad poll interval). All 36 findings graded via the
gadfly MCP (18 real / 18 false-positive).
gofmt clean, go vet quiet, go build + go test -race green.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Co-authored-by: Steve Dudenhoeffer <[email protected]>
Co-committed-by: Steve Dudenhoeffer <[email protected]>
After each review the binary POSTs the run + its heuristically-extracted findings to GADFLY_FINDINGS_URL (off unless set). Advisory: any error only goes to stderr — never touches stdout, the exit code, or the review. stdlib net/http only (no new deps). entrypoint.sh derives GADFLY_REPO/GADFLY_PR and passes through GADFLY_FINDINGS_URL/GADFLY_FINDINGS_TOKEN. Also renames store references from the old 'docket' name to 'gadfly-reports'.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Specialist lenses ran strictly sequentially within a model. Add a
GADFLY_LENS_CONCURRENCY knob (default 1 = unchanged) that overlaps the
independent per-lens review+recheck passes, so a model posts its
consolidated comment as soon as its lenses finish.
Per-provider configurable, mirroring GADFLY_PROVIDER_CONCURRENCY:
GADFLY_PROVIDER_LENS_CONCURRENCY takes a "provider=N,..." map keyed by
the same provider lanes (modelProvider() mirrors entrypoint's provider_of;
providerOverride() mirrors provider_cap). The override wins for the model's
lane, else the scalar default.
runSpecialists fans out via a bounded worker pool, order-preserving
(results written by index) and keeping each lens's own timeout/recheck.
repoFS is immutable + fresh-toolbox-per-pass, so lenses share no mutable
state (verified under -race). Docs/examples updated; dropped a duplicate
GADFLY_TIMEOUT_SECS README row.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Accept "foreman" in both resolveModel (GADFLY_BASE_URL) and endpointProvider
(GADFLY_ENDPOINT_*) switches, mapping to majordomo's ollama.Foreman() preset
(handles foreman's non-streaming/long-poll quirks). Unlike the HTTPS-only
LLM_* foreman:// DSN, the base URL is verbatim, so a plaintext http:// foreman
queue works. Tests + README provider table + endpoint-aliases example updated.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
A section that led with '**Blocking issues**' (no 'found') fell through to
unknown, so the consolidated header wrongly read 'No material issues found'
(seen live on gpt-oss). Now matches 'blocking issue'/'minor issue'/'no material
issue' and picks the earliest-appearing phrase (the lead verdict). + tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
A green check on a section reporting blocking issues was misleading; 🎯 signals
accuracy/on-target. Section verdict text already conveys pass/fail.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Five fixes, several surfaced by the live bake-off:
- PER-LENS TIMEOUT (critical): GADFLY_TIMEOUT_SECS now applies to EACH specialist
(own context), not shared across the suite. A slow model (e.g. a 35B local MLX)
was exhausting the whole 600s budget on lens 1, leaving the rest "step 0:
context deadline exceeded". Default lowered to 300s (per-lens). cmd/gadfly/main.go.
- ERRORED VERDICT: a lens whose review pass failed no longer counts as "clean".
Header shows "· ⚠️ N/M lens(es) errored" (or "Review incomplete — all lenses
errored"); the section reads "⚠️ could not complete". consolidate.go.
- PROVIDER LABEL: the comment header now shows the model's ACTUAL backend from the
spec ("m1pro/qwen3.6:35b-mlx" -> m1pro), not the global GADFLY_PROVIDER default
(was wrongly "ollama-cloud" for local models). scripts/run.sh.
- LENS FOCUS: base prompt no longer licenses "report anything serious"; each lens
stays in its lane, says "nothing in my area" rather than re-reporting another
lens's bug, with a one-line "Outside my lens:" escape hatch. The re-derive-
constants discipline is now lane-scoped, not "every lens". system-prompt.txt + specialists.go.
- RUN TIMING: run.sh posts a "⏳ Reviewing…" placeholder at model start and updates
it with "⏱️ reviewed in 1m 23s" on finish, for per-model comparison.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Two Phase-2 swarm upgrades:
- auto.go: GADFLY_SPECIALISTS=auto routes the review — a selector model
(GADFLY_SELECTOR_MODEL, else the review model) reads the changed files + PR
description and picks the smallest relevant lens set from the catalog, and may
propose ad-hoc lenses for gaps (e.g. migrations). Structured output via
majordomo.Generate[T]; capped + de-duped; falls back to the default suite.
- delegate.go: GADFLY_WORKER_MODEL adds a delegate_investigation tool so the
reviewer offloads mechanical legwork (trace callers, gather usages) to a cheap
worker sub-agent that returns an evidence-cited digest — the top model reasons
over summaries, not raw file dumps. Workers get an fs-only toolbox (no
sub-delegation). Unset = off.
resolveSpecialists now also returns the registry + an auto flag. Docs (README
Specialists + config table, CLAUDE.md, main.go header) + tests updated.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Replace the single generic review with a suite of focused specialists, each its
own review+recheck pass, merged into ONE comment (a collapsible section per lens,
led by the worst verdict; the optional `improvements` lens never escalates it).
- cmd/gadfly/specialists.go: built-in lenses + default suite (security, correctness,
maintainability, performance, error-handling) + opt-in (tests, docs, conventions,
improvements). Selection via GADFLY_SPECIALISTS (csv/"all"); custom defs via
GADFLY_SPECIALIST_<NAME> env and a repo .gadfly.yml (specialists + define).
Precedence: built-ins < file < env. Unknown names error but don't sink the run.
- cmd/gadfly/consolidate.go: verdict parse + one-comment render.
- main.go: loop specialists; per-lens failure is an inline notice, never fatal.
Default timeout bumped to 600s (suite runs sequentially).
- base system prompt trimmed to persona+tools+discipline+output; lens-specific
focus is appended per specialist (semantic re-derivation discipline kept in base).
- entrypoint default models -> single model (suite already gives breadth; cost ~=
specialists × models × 2). Adds gopkg.in/yaml.v3.
- docs/examples: README "Specialists" section, examples/.gadfly.yml, stub var,
CLAUDE.md architecture/config. Dynamic `auto` selection is the planned next step.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
majordomo's built-in LLM_* env DSNs are HTTPS-only (DSN.BaseURL forces https),
so they can't express a plaintext local Ollama. Add Gadfly-native env families
that register named providers/aliases with majordomo before resolution:
GADFLY_ENDPOINT_<NAME>="<provider>|<base-url>[|<key>]" # base URL verbatim (http ok)
GADFLY_ALIAS_<NAME>="<majordomo spec>" # plain alias / failover chain
Then reference them as "<name>/<model>" (or the bare alias) in GADFLY_MODEL(S).
<NAME> lowercases to the registry name, matching majordomo's LLM_* convention.
LLM_* DSNs still work (and are documented) for HTTPS endpoints. + unit tests,
README "Endpoint aliases via env vars", stub example.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Replace the hardcoded ollama.Cloud binding with majordomo's provider registry,
so Gadfly can target any backend majordomo supports without code changes.
- cmd/gadfly/model.go: resolveModel() — GADFLY_PROVIDER (default ollama-cloud)
prefixes bare model ids; GADFLY_MODEL may be a full provider/model spec, alias,
or failover chain (verbatim). GADFLY_BASE_URL constructs openai/ollama/anthropic/
google directly at a custom endpoint (OpenAI-compatible + local/remote Ollama).
GADFLY_API_KEY else the provider's standard env var. + buildSpec unit tests.
- run.sh: provider-aware key gate (local Ollama needs none); maps OLLAMA_CLOUD_API_KEY
-> OLLAMA_API_KEY; provider/base-url/key inherited by the binary. Gadfly-branded comment.
- entrypoint.sh: GADFLY_MODELS alias for OLLAMA_REVIEW_MODELS; provider passthrough.
- examples + README: Models & providers section. Upfront: only the Ollama paths
(local + OpenAI-compatible-against-Ollama) are tested; OpenAI/Anthropic/Google
are wired via majordomo but UNTESTED (no spend).
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Standalone, Docker-packaged extraction of the agentic PR reviewer that runs in
Gitea Actions: reads the checked-out repo with read-only tools (read_file/grep/
find_files/get_diff), verifies findings before reporting, two-pass review +
adversarial recheck, posts one labeled comment per model. Advisory only.
- cmd/gadfly: reviewer binary (majordomo + Ollama Cloud), zero deps beyond stdlib + majordomo
- entrypoint.sh: container brains — trigger gating, PR clone, model loop (logic out of YAML)
- Dockerfile: multi-stage; build-time module token never reaches the final image
- .gitea/workflows/build-image.yml: tag v* → build & push image
- examples/: ~15-line consumer stub
- system prompt genericized + hardened to re-derive constants/formulas (semantic bugs)
Vibe-coded with Claude Code; see README disclosure. Advisory, never blocks merge.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
steveandClaude Opus 5