fix: git slot suppresses SHA-only partial probe results (plan #660, source #659)

[btipling]

What

One-line guard in refreshGitStatusSlot: when the server probe returns a SHA-only value (starts with @, e.g. @abc1234), skip the update and keep the last honest branch@sha.

Source issue: #659
Plan: #660

Root cause

Two paths produce SHA-only partial results (branch missing, SHA present):

• HEAD rejection — rev-parse --abbrev-ref HEAD returns HEAD → probe strips it → @sha
• Transient git lock — agent exec holds lock → runGit returns '' → @sha

Both converge on refreshGitStatusSlot which unconditionally trusted any non-empty data.value.

Fix

if (data.value.startsWith('@')) return; // SHA-only → unreliable; keep last honest branch@sha

One new line between the non-empty check and setStatusSlot.

Files

File	Δ
lib/harnessChat.ts	+1 line — startsWith('@') guard
lib/harnessChat.test.ts	+70 lines — 7 new test cases
docs/harness-limits.md	+1 sentence — SHA-only suppression note on Git slot row

Test results

Suite	Result
npx vitest run lib/harnessChat.test.ts	✅ 150 passed
npx tsc --noEmit	✅ clean

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
invincible	Ignored		Aug 18, 2026 4:49am

[btipling]

Adversarial review — PR #661

Verdict: APPROVE
Repo: btipling/invincible
Scope: main ← plan/fix-git-slot-sha-only @ 73fb452 · 3 files · +72/−1 · git-slot SHA-only guard
Lenses run: L1 L3 L4 L5 L6 L8 L9 (skip: L2 — no API/secrets/runner; L7 — no host/config bind)
AGENTS.md read: yes · docs/feature-divide.md yes · SECURITY.md N/A

The mid-turn @sha clobber is real (formatGitStatusSlot with stripped/'' branch), the guard sits in the single funnel (refreshGitStatusSlot), empty-clear / 429 / network keep-last still work, and the 7 new tests hit the plan matrix. Layer is correct (host display policy, probe stays pure).

Findings

Sev	Lens	Finding	Break scenario	Refutation attempt	Confidence
Minor	L1	startsWith('@') rests on a false invariant. Plan #660 / plan-review: “git branch names cannot start with @.” git check-ref-format --branch '@foo' exits 0; git checkout -b '@hotfix' succeeds; --abbrev-ref HEAD returns @hotfix; formatGitStatusSlot then emits @hotfix@abc1234.	Operator (or agent) is on @hotfix / @1. Every cadence + on-demand refresh is dropped. Slot stays empty or frozen on the previous branch forever. The wire already distinguishes this: { git: { branch: "@hotfix", sha } } vs { git: { sha } } (no branch). Host types the JSON as { value?: string } and ignores git.branch.	Defender: “nobody names branches @….” Legal and I just created one. Fix: suppress when git.sha is set and git.branch is absent — not when the formatted string happens to start with @.	high
Minor	L9	Lasting detached HEAD is indistinguishable from a transient lock, so the bar keeps painting the previous branch name. Source #659 allowed showing @sha for a real detach; Goal 3 in the plan still says the cadence should. Decision A froze the last branch@sha and called it “honest.” After checkout --detach / rebase / bisect it is a lie.	Detach, commit, keep working. Slot still says main@old. Operator thinks they are on main.	Defender: “plan-review accepted the tradeoff.” Accepted ≠ honest. Mid-turn-only suppression (plan option B) or “two consecutive cadence SHA-only ticks → accept @sha” would keep the flicker fix without a stale branch. Not a merge block; do not document the leftover as “honest.”	high
Nit	L6	New 429 / network cases duplicate the existing keep-last test in the same describe. No case for dirty SHA-only (@abc*, which formatGitStatusSlot can emit) or for @hotfix@sha pass-through.	A later edit to the 429 short-circuit is still green because the @ guard would also keep-last.	Rename / drop the dupes; add the two missing shapes if the discriminator stays string-based.	high

Residual risk

No unit-test workflow ran on this SHA (repo has no vitest CI; only Vercel ignore). I did not re-run vitest in this workspace. Detached-at-boot stays empty (plan-accepted, same as non-git). refreshGitStatusSlot’s fail-soft comment was not updated to mention SHA-only.

Merge guidance

APPROVE — the reported mid-turn @sha flash is fixed in the right layer. Ship if the Minors stay residual. Prefer keying the guard on structured git.branch / git.sha from the same 200 body before a follow-up if @-prefixed branch names or lasting detach matter.

What was not attacked

Live /harness mid-exec lock, real checkout --detach on a bound sandbox, prod Gateway, DO runner.

[btipling]

Adversarial review — PR #661 (verdict correction)

Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/fix-git-slot-sha-only @ 73fb452 · 3 files · +72/−1 · git-slot SHA-only guard
Lenses run: L1 L3 L4 L5 L6 L8 L9 (skip: L2 — no API/secrets/runner; L7 — no host/config bind)
AGENTS.md read: yes · docs/feature-divide.md yes · SECURITY.md N/A

Supersedes the earlier comment on this PR that said APPROVE. That is not a skill verdict. Official scale is BLOCK / CONCERNS / PASS WITH NOTES / INCOMPLETE. Findings unchanged; two Minors + one Nit → PASS WITH NOTES. merge-pr can treat this head as review-satisfied.

The mid-turn @sha clobber is real (formatGitStatusSlot with stripped/'' branch), the guard sits in the single funnel (refreshGitStatusSlot), empty-clear / 429 / network keep-last still work, and the 7 new tests hit the plan matrix. Layer is correct (host display policy, probe stays pure).

Findings

Sev	Lens	Finding	Break scenario	Refutation attempt	Confidence
Minor	L1	startsWith('@') rests on a false invariant. Plan #660 / plan-review: “git branch names cannot start with @.” git check-ref-format --branch '@foo' exits 0; git checkout -b '@hotfix' succeeds; --abbrev-ref HEAD returns @hotfix; formatGitStatusSlot then emits @hotfix@abc1234.	Operator (or agent) is on @hotfix / @1. Every cadence + on-demand refresh is dropped. Slot stays empty or frozen on the previous branch forever. The wire already distinguishes this: { git: { branch: "@hotfix", sha } } vs { git: { sha } } (no branch). Host types the JSON as { value?: string } and ignores git.branch.	Defender: “nobody names branches @….” Legal and I just created one. Fix: suppress when git.sha is set and git.branch is absent — not when the formatted string happens to start with @.	high
Minor	L9	Lasting detached HEAD is indistinguishable from a transient lock, so the bar keeps painting the previous branch name. Source #659 allowed showing @sha for a real detach; Goal 3 in the plan still says the cadence should. Decision A froze the last branch@sha and called it “honest.” After checkout --detach / rebase / bisect it is a lie.	Detach, commit, keep working. Slot still says main@old. Operator thinks they are on main.	Defender: “plan-review accepted the tradeoff.” Accepted ≠ honest. Mid-turn-only suppression (plan option B) or “two consecutive cadence SHA-only ticks → accept @sha” would keep the flicker fix without a stale branch. Not a merge block; do not document the leftover as “honest.”	high
Nit	L6	New 429 / network cases duplicate the existing keep-last test in the same describe. No case for dirty SHA-only (@abc*, which formatGitStatusSlot can emit) or for @hotfix@sha pass-through.	A later edit to the 429 short-circuit is still green because the @ guard would also keep-last.	Rename / drop the dupes; add the two missing shapes if the discriminator stays string-based.	high

Residual risk

No unit-test workflow ran on this SHA (repo has no vitest CI; only Vercel ignore). I did not re-run vitest in this workspace. Detached-at-boot stays empty (plan-accepted, same as non-git). refreshGitStatusSlot’s fail-soft comment was not updated to mention SHA-only.

Merge guidance

PASS WITH NOTES — the reported mid-turn @sha flash is fixed in the right layer. Safe to merge from this attack. Prefer keying the guard on structured git.branch / git.sha from the same 200 body in a follow-up if @-prefixed branch names or lasting detach matter.

What was not attacked

Live /harness mid-exec lock, real checkout --detach on a bound sandbox, prod Gateway, DO runner.

[btipling]

Adversarial review — PR #661 (round 2)

Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/fix-git-slot-sha-only @ 9c1ad2f · 3 files · SHA-only git-slot guard
Lenses run: L1 L3 L4 L5 L6 L8 L9 (skip: L2 — no API/secrets/runner; L7 — no host/config bind)
AGENTS.md read: yes · docs/feature-divide.md yes · SECURITY.md N/A

Supersedes reviews on 73fb452 (unofficial APPROVE, then PASS WITH NOTES). New head keys the guard on structured git.sha && !git.branch instead of value.startsWith('@').

The mid-turn @sha clobber is still fixed in the single funnel. Layer is still correct (host display policy; probe stays pure). @hotfix@sha now passes through. Dirty SHA-only (git.sha + dirty, no branch) is covered.

Findings

Sev	Lens	Finding	Break scenario	Refutation attempt	Confidence
Minor	L9	Lasting detached HEAD is still the same wire shape as a transient lock (sha set, branch absent), so the bar keeps the previous branch name. Docs and the new comment still call that the “last honest branch@sha.” After checkout --detach / rebase / bisect it is a stale name.	Detach, commit, keep working. Slot still says main@old. Operator thinks they are on main.	Defender: “plan Decision A + we named detached HEAD in the docs.” Named ≠ honest. Mid-turn-only suppression or “two cadence SHA-only ticks → accept @sha” would keep the flicker fix. Not a merge block.	high
Nit	L8	refreshGitStatusSlot’s fail-soft bullet list (network / empty / ellipsize) was not updated for the new SHA-only return.	Next editor treats “any non-empty value” as the whole policy and deletes the structured guard.	Add one bullet: sha-present + branch-absent → keep last.	high

Prior findings (round 1 @ 73fb452) — disposition

Prior	Status
Minor L1 startsWith('@') false invariant / @hotfix	Fixed. Guard is data.git?.sha && !data.git?.branch. New test paints @hotfix@abc1234.
Minor L9 stale branch after detach	Still open (same policy).
Nit L6 duplicate 429/network + missing @sha* / @hotfix	Mostly fixed. Dirty SHA-only and @hotfix tests added. 429/network cases still overlap the older keep-last test; not re-raised.

Residual risk

If a 200 body has value: '@abc' and omits git, the guard does not fire (git?.sha is missing) and @abc still paints. The live route always includes git; this is fail-open only on a hand-rolled / future-thinner payload. No vitest CI on this SHA (repo has none; only Vercel ignore). I did not re-run vitest here. Detached-at-boot stays empty (plan-accepted).

Merge guidance

PASS WITH NOTES — review-satisfied for merge-pr. The reported mid-turn @sha flash is fixed; the structured discriminator closes the @hotfix hole. Lasting-detach staleness stays residual.

What was not attacked

Live /harness mid-exec lock, real checkout --detach on a bound sandbox, prod Gateway, DO runner.