Skip to content

Forgejo/Gitea forge parity: pr-search, pr-diff, and the pr-exists hang - #19

Merged
pseudoseed merged 23 commits into
mainfrom
builder/pir-12
Aug 21, 2026
Merged

Forgejo/Gitea forge parity: pr-search, pr-diff, and the pr-exists hang#19
pseudoseed merged 23 commits into
mainfrom
builder/pir-12

Conversation

@pseudoseed

@pseudoseed pseudoseed commented Aug 21, 2026 •
edited
Loading

Copy link
Copy Markdown
Owner

PIR Review: Forgejo/Gitea forge parity — pr-search, pr-diff, and the pr-exists "hang"

Fixes #12

Summary

A Forgejo repository can now run on the bare forge.provider: gitea preset with no per-concept overrides and no gh shim. pr-search and pr-diff are implemented for gitea and no longer disabled; pr-exists stopped enumerating pull requests and answers in one request instead of about seventeen minutes; and recently-merged, which had been silently returning nothing on every afx status for months, went from about twenty-six minutes to roughly one second.

The reported bug was a hang. It was not one — it was a ~17-minute loop, and finding that out changed the fix from "add a timeout" to "stop enumerating."

Files Changed

  • packages/codev/scripts/forge/gitea/_lib.sh (+245 / -0) — timeout, bounded-concurrency fetch, repo/default-branch resolution, paged-walk deadline
  • packages/codev/scripts/forge/gitea/pr-exists.sh (+104 / -0) — rewritten around the base/head lookup
  • packages/codev/scripts/forge/gitea/pr-search.sh (+199 / -0) — new
  • packages/codev/scripts/forge/gitea/pr-diff.sh (+76 / -0) — new
  • packages/codev/scripts/forge/gitea/recently-merged.sh (+174 / -0) — rewritten around the cheap issues index
  • packages/codev/scripts/forge/gitea/pr-list.sh (+14 / -0) — truncation is no longer swallowed by a pipe
  • packages/codev/scripts/forge/gitea/issue-view.sh (+1 / -0) — # forge-executable: tea
  • packages/codev/src/lib/forge.ts (+57 / -0) — gitea preset enables both concepts; shared "concept unavailable" reporting
  • packages/codev/src/lib/forge-contracts.ts (+17 / -0) — PrSearchItem gains state, baseRefName, title, url
  • packages/codev/src/lib/github.ts (+23 / -0) — a preset-disabled on-it-timestamps warns instead of returning empty
  • packages/codev/src/lib/team-github.ts (+19 / -0) — names the provider instead of "returned no data"
  • packages/codev/src/commands/porch/checks.ts (+17 / -0) — null from pr-exists is "could not answer", not "no PR"
  • packages/codev/src/agent-farm/commands/spawn-worktree.ts (+12 / -0) — the collision query says is:open
  • packages/codev/src/__tests__/pir-12-gitea-pr-concepts.test.ts (+~670 / -0) — new
  • packages/codev/src/commands/porch/__tests__/pir-12-pr-exists-null-vs-false.test.ts (+77 / -0) — new
  • packages/codev/src/__tests__/bugfix-1137-gitea-tea-api.test.ts (+110 / -0) — fixtures follow the endpoints
  • packages/codev/src/agent-farm/__tests__/spawn-worktree.test.ts (+18 / -0)
  • .claude/skills/forge/SKILL.md, .codex/skills/forge/SKILL.md (+50 each) — kept byte-identical
  • codev/resources/arch.md (+11 / -0), codev/resources/lessons-learned.md (+6 / -0), codev/resources/lessons-critical.md (+1 / -1)
  • codev/plans/12-forgejo-gitea-forge-parity.md (+217 / -0), codev/state/pir-12_thread.md (+111 / -0)

24 files changed, 2108 insertions, 75 deletions.

Commits

The measurements that decided the design

Everything here follows from timings taken against a live Forgejo 15.x (git.pseudoseed.com/pseudoseed/entriq, 1599 PRs) with tea 0.14.2, not from reasoning about the API.

Request Time
GET /repos/{o}/{r}/issues/1 0.59 s
GET .../pulls?state=all&limit=1 0.78 s
GET .../pulls?state=all&limit=50 32.8 s
GET .../pulls?state=closed&limit=50 48.1 s
GET .../issues?type=pulls&state=all&limit=50 1.75 s
GET .../pulls/{base}/{head} 0.17–1.2 s
GET .../pulls/{n}.diff 0.30 s

The /pulls list is priced per returned PR object, not per request — roughly 0.65 s each, because Gitea materialises head and base commit info for every row. pr-exists walked state=all at 50 per page, so it cost 32 pages × ~33 s ≈ 17 minutes; recently-merged walked state=closed at 48 s a page ≈ 26 minutes. Raising the page size cannot help a per-item cost, and neither can a cache that has to be filled once. The fix had to stop enumerating.

Two properties of Gitea made that possible, and one of them contradicts a caveat the old code documented as unfixable:

  • GET /pulls/{base}/{head} answers "is there a PR from this branch" in one request, 404s when there is none, and takes slashes in the head branch unescaped.
  • head.label survives branch deletion. The old scan matched .head.ref, and its comment correctly noted that Gitea rewrites that to refs/pull/N/head once a merged PR's branch is deleted — concluding that a merged PR could not be found by branch name. But .head.label keeps the original name and the base/head endpoint matches on the stored head branch. Verified: PR 3869, merged with its branch gone, reports head.ref = refs/pull/3869/head and head.label = builder/aspir-3860, and pulls/main/builder/aspir-3860 returns it. So the new implementation is not only ~900× faster, it answers a question the old one could not.

recently-merged uses a different lever: the issues?type=pulls index is nineteen times cheaper than the /pulls view of the same rows because it skips that commit materialisation, and it accepts a server-side since filter. It carries everything MergedPrItem needs except the head branch, which is then fetched per match, eight at a time.

The codev doctor fix, found during the acceptance run

Separate from the above and worth its own heading, because it silenced the one diagnostic that should have caught this class of problem.

extractExecutable reads a script's first substantive line to decide what must be on PATH. For the gitea preset it was answering echo for issue-view and case for pr-list — so codev doctor on a Forgejo repo told the user to install echo, and a genuinely missing tea went unreported on exactly the repositories that need it. This is the defect class cluesmith#1455 exists to close, and the remedy it established is the # forge-executable: declaration; two scripts were missing one. The test now asserts it across the entire preset rather than concept by concept, so a new gitea script cannot reintroduce it.

What upstream cluesmith#1331 contributed, and what its review contributed

cluesmith#1331 (open, unmerged; no PR against it in this fork) fixes pr-search to span all PR states, because gh pr list --search defaults to open-only and post-merge consult --type pr therefore fails with "No PR found for branch". The gitea implementation defaults to all states for that reason.

Its review turned out to matter more than its diff. Making pr-search all-states breaks spawn-worktree.ts, which queries in:body #N and leaned on the old open-only default to mean "open PRs". Under an all-states search that becomes "did this issue ever have a PR", and every re-spawn, every follow-up to a partial fix, and every retry after a closed PR aborts with a factually wrong "Found N open PR(s)". The call site now says is:open explicitly, and the gitea script honours is: qualifiers so that saying it works. The review's second point — that PrSearchItem gave callers no state with which to defend themselves — is addressed by adding state to the contract and ordering results open-first.

github/pr-search.sh and gitlab/pr-search.sh are deliberately not touched here. They carry the same cluesmith#759 bug, but fixing them is upstream cluesmith#1331's diff and would conflict on every future sync. The architect is filing that as an adopt-the-PR job, the way cluesmith#1146 and cluesmith#1458 were handled.

The truncation contract

recently-merged bounds itself three ways — a default 7-day window when CODEV_SINCE_DATE is absent (announced on stderr; never "all time", which is the 26-minute walk), a 300-PR ceiling, and the paged walk's wall-clock deadline. When a bound bites it exits 3 with empty stdout.

That emptiness is the point. A short list and a truncated list are indistinguishable once printed, so [] with status 0 means "nothing merged" and status 3 means "I stopped looking", and stderr names which bound bit. This is now the documented exit-status contract for forge concepts generally.

Test Results

  • npm run build: ✓ pass

  • npm test: 5498 passed, 2 failed — both pre-existing and unrelated (see Flaky Tests). 45 tests are new across two files (pir-12-gitea-pr-concepts.test.ts, pir-12-pr-exists-null-vs-false.test.ts), plus one behavioural spawn assertion and the reworked bugfix-1137 cases.

  • Live verification against git.pseudoseed.com/pseudoseed/entriq with all three overrides deleted, driven through the real dispatcher (config load → preset → env → script → JSON parse), not by invoking scripts by hand:

    Call Time Result
    pr-exists open branch 986 ms true
    pr-exists merged branch, deleted 686 ms true
    pr-exists absent branch 482 ms false
    pr-search head:<branch> 1079 ms PR 3855 with head + base refs
    pr-search <issue number> 1240 ms PR 3869, state merged
    pr-search in:body #3860 is:open 620 ms [] — spawn not blocked
    pr-diff name-only / full 578 / 354 ms 6 paths / 30 KB diff
    issue-view 607 ms correct JSON
    pr-list 509 ms 1 open PR
    recently-merged 24 h 1082 ms 6 PRs, all head branches resolved
    user-identity 459 ms pseudoseed
    team-activity, on-it-timestamps — still resolve as disabled

    entriq's config was then restored and diff-verified byte-identical to both the pre-test copy and the architect's 11:20 backup. It runs the globally installed codev 3.3.1, whose preset still disables pr-search/pr-diff, so its overrides remain load-bearing until this ships. Delete them after this merges and entriq updates.

⚠️ This had TWO of THREE review lanes — Codex never ran

Read this before trusting the review depth.

Lane Verdict Notes
codex (gpt-5.6-sol) NEVER RAN Provider usage quota exhausted. Two attempts on 2026-08-21, 19:06:05Z and 19:06:35Z, each refused in ~6 s before any model work began. Quota restores 2026-08-27. No codex findings exist for this change.
gemini (agy) APPROVE, HIGH No issues raised.
claude (opus-5) APPROVE, HIGH Four non-blocking findings, all real, all fixed — see below.

Verbatim provider message, both attempts:

You've hit your usage limit. Upgrade to Pro (https://chatgpt.com/explore/pro), visit https://chatgpt.com/codex/settings/usage to purchase more credits or try again at Aug 27th, 2026 4:01 PM.

The codex lane was skipped on explicit architect instruction rather than hold a fork-local change for six days. The absence is recorded as a NOT-RUN file at codev/projects/12-forgejo-gitea-forge-parity-imp/12-review-iter1-codex.txt carrying VERDICT: SKIPPED and CONFIDENCE: NONE, so it cannot be misread as a review that happened — but that path is gitignored (.gitignore:65), which is why the coverage is stated here too. The same quota blocked codex on #2, #4 and #11 earlier the same day.

Porch's own gate summary says otherwise, and it is wrong. porch next 12 reports codex: COMMENT and prints "All reviewers approved!". parseVerdict (porch/verdict.ts:41-47) recognises only APPROVE, REQUEST_CHANGES and COMMENT; VERDICT: SKIPPED matches none of them, falls through to the "no valid VERDICT line found — treat as COMMENT" default, and allApproved counts COMMENT as approval. So a lane that never ran is indistinguishable from one that approved, in the exact summary a human reads when deciding to merge.

There is a second, independent route to the same false statement, found by the architect when verifying this in the source: allApprove also returns true when reviews.length === 0, so zero reviewers approves as well. And the fall-through's own comment states the bug aloud — "no valid VERDICT line found but the consult ran" — because the vocabulary was built for a reviewer that ran and emitted something unparseable. A reviewer that never ran cannot be expressed in it at all, which is why widening a default would not have been enough.

That is this PR's own lesson pointed back at the tooling — a sixth arrival at it in one day, and the one with the largest blast radius, since it affects every project that records a skip. #2, #4 and #11 all merged with porch reporting unanimous approval while codex had read nothing; those merges were sound because the human approving each gate knew the coverage was two of three and each PR body said so, but that was the human channel working despite the tool. It is porch behaviour rather than anything in this diff, so it is not fixed here — tracked as #20. Read the table above, not porch's summary line.

Weigh this specifically. The absent lane is the one that most often catches shell-quoting and POSIX-portability defects, and this diff is five POSIX sh scripts, a hand-rolled process watchdog, and a pile of jq. That is close to the worst pairing of which lane is missing against what the change is made of. Two of the three bugs found during implementation were exactly that class, and both were caught by tests rather than by reading.

What the Claude lane found, and what changed because of it

All four findings were reproduced before being fixed; none were argued down.

  1. Two gitea_api_error checks were unreachable (pr-diff.sh name-only, recently-merged.sh). Gitea answers a 404 with a JSON object, which reached tea_api_paged's jq -s 'add' first — and an object cannot be added to an array, so the walk died on jq: error … array ([]) and object ({...}) cannot be added before the script's own classification ran. Reproduced against live Forgejo exactly as reported. The exit status was non-zero either way, so no wrong answer was ever returned; what was lost was the sentence naming the missing PR. tea_api_paged now classifies before accumulating and returns a distinct status 4 with the body on stdout, so the caller's message finally fires. Both paths now tested — the full-diff 404 was covered and the name-only one was not, which is why only one of them was broken.
  2. The plan's Test Plan called for a byte-identity test over the two SKILL.md twins and it was never written. The files were identical; nothing stopped them drifting. Now pinned.
  3. checks.ts's null-vs-false change had no test. It is a real behaviour change — null from the concept means "could not answer", not "no PR exists" — and it mattered concretely, since before this PR the gitea script blew the 30 s ceiling on every run, making null the normal outcome on that provider. Three tests added in a separate file, because mocking the forge layer is something checks.test.ts deliberately avoids.
  4. gitea_timeout nits, both fixed. The marker file's path was derived from the output file's ("$_tf.fired"), which mktemp does not reserve — a predictable name in a world-writable tmpdir that anyone could pre-create to force every call to report a timeout. Both files now live in a private mktemp -d. And the watchdog claimed the timeout unconditionally, so a command finishing in the same instant the deadline passed was reported as timed out, discarding a good answer; it now checks kill -0 first. That narrows the window rather than closing it — a lock would be needed for that — and a false timeout is a retryable error rather than a wrong answer.

Architecture Updates

COLD — codev/resources/arch.md, § Integration Points → Forge Concept Commands. Four additions, all current-state reference rather than changelog:

  1. Provider presets are on-disk scripts under packages/codev/scripts/forge/<provider>/, single-source with no skeleton mirror, and null in a preset is a deliberate refusal rather than "no script".
  2. A preset-disabled concept is invisible to forgeConfig lookups — forgeConfig?.['x'] reads user config only. Availability must be asked of getForgeCommand/isConceptDisabled. This is the system-shape surprise that made on-it-timestamps return an empty map silently on every gitea repo.
  3. The exit-status contract, including 3-with-empty-stdout, and that null from executeForgeCommand means "could not answer".
  4. The two verified Forgejo properties: per-object pricing on /pulls, and head.label surviving branch deletion. Both filed here rather than in lessons-learned because they are properties of a system, not general engineering advice — the "looks like X but is actually Y" routing.

Nothing was promoted to arch-critical.md: all of it is reference detail that matters only when writing a forge concept.

Lessons Learned Updates

HOT — codev/resources/lessons-critical.md, one promotion with displacement, since the file was at its cap of ten:

A truncated result is indistinguishable from a complete one once emitted — give "I stopped early" its own signal and emit nothing, never a partial answer that reads as whole.

It earns the hot slot because the same principle has now been arrived at independently three times in this codebase — afx send's render gate, #13's log extraction, and this PR's exit-3 contract — which is the signal that it should be changing decisions before they are made rather than being rediscovered.

Displaced into lessons-learned.md § Architecture: "Model permissions as roles/capabilities, not booleans." Still true and still worth reading; it is the narrowest of the ten, applying only when designing a permission system, where the promoted rule applies to any list, log, render, or fetch that can stop early. Reviewers should push back if they disagree with that trade — it is the judgement call in this diff that is least supported by evidence.

COLD — codev/resources/lessons-learned.md § Debugging and Root Cause Analysis, three entries:

  1. Killing a process is not unblocking the caller, and a zero exit is not success. Both halves were bugs in this PR's own timeout helper (below).
  2. Measure whether a remote endpoint is priced per request or per returned item before optimising around it. One limit=1 vs limit=50 timing tells you which regime you are in and takes a minute; against a per-item cost, every page-size instinct is useless.
  3. A hang reported from a shell is not always the hang your code has in production. In-process, pr-exists was being killed at 30 s and returning null, which the caller rendered as "no PR exists" — same root cause, different bug, and only one of them was the one reported.

Things to Look At During PR Review

  • gitea_timeout in _lib.sh is the subtlest thing here, and it was wrong twice. First it killed the command and left the caller blocked anyway: every caller runs it inside $(...), and a grandchild kept the write end of that pipe open, so the timeout message printed at 3 s and the script was still hung two minutes later. The command now writes to a temp file rather than the caller's pipe, and the watchdog's own stdout goes to /dev/null for the same reason. Then it classified timeouts by exit status (143/137), which a wrapper using operand-less wait defeats — POSIX makes that always return 0, so a process killed by SIGTERM reported success with an empty body and the caller diagnosed an unreadable repository instead. The watchdog now records that it fired and nothing is inferred. Both failure modes have tests.
  • The base-branch limitation in pr-exists / pr-search head:. The endpoint needs a base; it uses the repository default unless CODEV_PR_BASE is set, so a PR against an integration branch needs that variable. A miss prints false, which fails the porch pr_exists gate loudly. Falling back to a list scan on the 404 path was rejected deliberately — it would reintroduce the ~17-minute walk on the failure path, where it would be least expected.
  • pr-search's query grammar is a parser, not a pass-through, covering exactly the five query strings this codebase builds. A sixth would return [] with a stderr note. The five and their call sites are tabulated at the top of the script.
  • bugfix-1137's fixtures moved. Its guarantee is unchanged and still asserted — reads go through tea api, never tea pulls list, and a paged read is not silently truncated — but pr-exists now pages nothing at all, which is the strongest form of that. The fake tea no longer serves any pulls?state=… fixture, so a reintroduced scan fails there with "no fixture for".
  • Script assertions are anchored to command lines, not toContain. Fix #759: pr-search includes all PR states so merged PRs are found cluesmith/codev#1331's review caught assertions that passed against the explanatory comment quoting the flag under test, staying green with the flag deleted from the command. The comments here quote the very endpoints being pinned, so the same trap was live.
  • The hot-tier displacement described above.

How to Test Locally

  • View diff: VSCode sidebar → right-click builder pir-12 → Review Diff
  • Run dev: afx dev pir-12
  • What to verify, against any Forgejo repo with tea authenticated:
    • CODEV_BRANCH_NAME=<open-branch> sh packages/codev/scripts/forge/gitea/pr-exists.sh returns true in about a second, and issues exactly two requests
    • the same for a branch whose PR is merged and whose branch was deleted — this is the case the old code could not answer
    • a branch with no PR returns false; a bad CODEV_REPO errors rather than returning false
    • CODEV_SEARCH_QUERY='in:body #<n> is:open' returns [] for an issue whose PR is merged, i.e. afx spawn is not blocked
    • CODEV_SINCE_DATE unset on recently-merged.sh announces a 7-day window rather than walking everything; CODEV_FORGE_MERGED_MAX=2 over a busy window exits 3 with empty stdout
    • codev doctor reports tea for every enabled gitea concept and disabled for team-activity / on-it-timestamps

Flaky Tests

None skipped or annotated. Two pre-existing failures were left untouched (which of them fires varies per run — they are contention-sensitive, and a third from the same file appeared in an earlier run):

  • packages/codev/src/__tests__/spec-1280-measurement-instrument.test.ts — emits byte-identical output twice at the same commit, reports the same total under a C locale as under UTF-8, and PHASE_ITERS is a linear comparison constant. Each invokes scripts/measure-prompt-surface.sh twice against a 60 s budget, and that script takes ~31 s per invocation on this machine. Proven pre-existing by running the same test against the unmodified main checkout, where it fails identically at 77 s; the architect independently reproduced it there past 300 s. The script costs the same against either root (31.3 s vs 33.0 s), so this is machine speed, not diff content.

packages/codev/src/terminal/__tests__/session-manager.test.ts also drops one timing-sensitive stderr-tail case under full-suite load; the file passes 91/91 in isolation, and a different case from it failed on an earlier run, which is the signature of contention rather than a defect.

This is further evidence for #8 — porch's check timeout is a hardcoded 300 s with no override key, and entriq's config records the same class of problem at 460 s quiet / 859 s contended.

pseudoseed and others added 23 commits August 21, 2026 11:29
@pseudoseed
chore(porch): 12 init pir
0133cb4
@pseudoseed @claude
[PIR #12] Plan draft: gitea pr-search, pr-diff, and the pr-exists hang
28e764f 
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@pseudoseed
chore(porch): 12 plan-approval gate-requested
362cab3
@pseudoseed
chore(porch): 12 plan-approval gate-approved
102dd14
@pseudoseed
chore(porch): 12 implement phase-transition
594bd56
@pseudoseed @claude
[PIR #12] fix(gitea): answer pr-exists in one request instead of ~17 …
63c9807 
…minutes

Forgejo charges the /pulls list endpoint per RETURNED PR OBJECT — measured
0.78s at limit=1 and 32.8s at limit=50 against Forgejo 15.x. Walking
state=all over a 1599-PR repo therefore cost ~17 minutes for one yes/no
question, which read as a hang.

GET pulls/{base}/{head} answers it in ~1.2s, and answers it better: the old
scan matched .head.ref, which Gitea rewrites to refs/pull/N/head once a
merged PR's branch is deleted, so merged PRs could not be found by branch
name at all. .head.label keeps the branch name and this endpoint matches on
it.

Also adds pr-search and pr-diff for gitea, replaces recently-merged's
26-minute walk with the cheap issues index plus bounded concurrent head
lookups, and gives every gitea script a wall-clock timeout so a future hang
surfaces as a named error rather than a stuck phase.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@pseudoseed @claude
[PIR #12] fix: enable the gitea concepts, and stop degrading silently
d066687 
- forge.ts: pr-search and pr-diff are no longer disabled for gitea.
- spawn-worktree: the collision query says 'is:open' instead of leaning on
  pr-search's old open-only default. Without this, an all-states search
  aborts every re-spawn of an issue that ever had a merged PR with a
  factually wrong 'Found N open PR(s)' (upstream cluesmith#1331's review).
- github/team-github: a concept disabled by a PROVIDER PRESET was invisible
  to forgeConfig lookups, so on-it-timestamps returned an empty map with
  nothing on stderr. Both callers now name the provider and the consequence.
- checks.ts: a null from the pr-exists concept is reported as 'could not
  answer', not as 'no PR exists'.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@pseudoseed @claude
[PIR #12] test+docs: pin the endpoints, the head.label behaviour, and…
bbc4ebe 
… is:open

Assertions are anchored to command lines rather than matched with toContain,
so the explanatory comments — which quote the very strings under test —
cannot keep a test green after the code it pins is removed (cluesmith#1331's review
caught exactly that).

bugfix-1137's fixtures move with the endpoints they mock. Its guarantee is
unchanged and still asserted: reads go through tea api, never tea pulls
list, and a paged read is not silently truncated.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@pseudoseed @claude
[PIR #12] fix(doctor): report tea for every gitea concept, not echo o…
64cb220 
…r case

Found by the acceptance run against entriq. extractExecutable reads a
script's first substantive line, which answered 'echo' for issue-view and
'case' for pr-list — so `codev doctor` on a Forgejo repo told the user to
install `echo`, and a genuinely missing `tea` went unreported. That is the
cluesmith#1455 defect class, and the '# forge-executable:' declaration is its remedy.

The test now asserts it across the whole preset rather than concept by
concept, so a new gitea script cannot reintroduce it.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@pseudoseed @claude
[PIR #12] docs: builder thread for the implement phase
e018fee 
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@pseudoseed
chore(porch): 12 dev-approval gate-requested
bb40271
@pseudoseed
chore(porch): 12 dev-approval gate-approved
6da756b
@pseudoseed
chore(porch): 12 review phase-transition
13c8cbe
@pseudoseed @claude
[PIR #12] Review + retrospective
c359c60 
Routes the findings by tier: the Forgejo per-object pricing and the
head.label survival go to arch.md as system-shape properties, the
partial-result rule is promoted to the hot lessons tier (displacing the
permissions/booleans lesson into cold), and three debugging lessons land in
lessons-learned.md.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@pseudoseed
chore(porch): 12 record PR #19
2e784d2
@pseudoseed
chore(porch): 12 review build-complete
64c50fe
@pseudoseed @claude
[PIR #12] docs: builder thread for the review phase
f541a94 
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@pseudoseed @claude
[PIR #12] fix: address review round 1 (claude, 4 non-blocking findings)
ed297bf 
1. A Gitea error body is a JSON object, which reached tea_api_paged's
   `jq -s 'add'` before any caller could classify it — an object cannot be
   added to an array, so the walk died on a raw jq diagnostic and the
   script's own 'PR #N not found' message was unreachable. Reproduced
   against live Forgejo. Classify before accumulating; status 4 carries the
   body out to the caller. Never returned a wrong answer, only a worse
   message — and the name-only path was the untested one, which is why it
   was the broken one.
2. The plan called for a SKILL.md byte-identity test that was never written.
3. checks.ts's null-vs-false change had no test; three added.
4. gitea_timeout: the marker's path was predictable in a world-writable
   tmpdir (both files now in a private mktemp -d), and the watchdog claimed
   a timeout for a command that finished in the same instant — it checks
   kill -0 first now.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@pseudoseed
chore(porch): 12 pr gate-requested
136fb77
@pseudoseed @claude
[PIR #12] docs: porch reports a SKIPPED lane as an approval
269d2d7 
parseVerdict knows APPROVE/REQUEST_CHANGES/COMMENT and nothing else, so the
VERDICT: SKIPPED that the NOT-RUN convention writes falls through to the
COMMENT default, and allApproved counts COMMENT as approval. porch therefore
printed 'All reviewers approved!' for a codex lane that never ran.

Not fixed here — it is porch behaviour affecting every project that records a
skip, not something in this diff. Recorded in the review so a human reading
the gate summary is not misled, and in lessons-learned so the next person
writing an absence-recording convention checks who reads it.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@pseudoseed
chore(porch): 12 pr gate-approved
8d4e181
@pseudoseed
chore(porch): 12 protocol complete
bbd64aa
@pseudoseed @claude
[PIR #12] docs: the porch verdict hole has two routes, not one (#20)
5573cbd 
Architect verified in source: besides parseVerdict falling through to
COMMENT, allApprove separately returns true when reviews.length === 0, so
zero reviewers approves too. And the fall-through's own comment says 'but
the consult ran' — the vocabulary was built for a reviewer that ran and
emitted something unparseable, so a reviewer that never ran is not merely
unhandled, it is inexpressible. Widening a default would not have fixed it.

Also records that #2, #4 and #11 merged under the same false summary.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@pseudoseed
pseudoseed merged commit c70ce95 into main Aug 21, 2026
7 checks passed
pseudoseed added a commit that referenced this pull request Aug 21, 2026
@pseudoseed
chore(porch): 12 PR #19 merged
19ef2f0
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

Forgejo/Gitea forge parity: implement pr-search and pr-diff, fix the pr-exists hang

1 participant

@pseudoseed