fix(agent): pre-action verifier and belief state run nowhere — AgentLoop has no production caller

Child of #13587.

Problem

The pre-action adversarial verifier and the belief-state tracker both run in
production nowhere. Their only caller is AgentLoop, which has no production
instantiation — the live path is the LangGraph StateGraph in
chat_workflow/graph.py plus chat_workflow/tool_handler.py.

Evidence

grep -rn "from agent_loop.loop import\|AgentLoop(" across the repo (excluding
worktrees and venv) returns only test files:

autobot-backend/agent_loop/test_loop_repetition.py:22,44,212
autobot-backend/agent_loop/test_approval_workflow.py:25,53
autobot-backend/agent_loop/spill_loop_wiring_test.py:183,185
autobot_shared/tool_catalogue_test.py:76,194

Both components are reachable only from that dead entry point:

PreActionVerifier — sole non-test caller is agent_loop/loop.py:191-192,1668.
pre_action_verifier_enabled defaults to True (agent_loop/types.py:266),
so it reads as an active guard while executing on no request.
BeliefState / BeliefStateUpdater — sole non-test caller is
agent_loop/loop.py:28,186,644,699,1350,1385. belief_state_enabled defaults
to False (agent_loop/types.py:258): off and unreachable.

Correction to this umbrella's own premise

#13587's "Already covered — audited, confirmed present, NOT filed" section lists:

Pre-action adversarial verifier — agent_loop/pre_action_verifier.py,
verdict surfaced to the human at the approval gate (agent_loop/loop.py:1600,1676).
No external equivalent.

The citation is to a module with no production caller. The capability is real and
well built; it is not covered, because it never executes. The same correction
applies to the "Durable run checkpointing" entry, which also cites agent_loop/loop.py.

By contrast, the entries citing chat_workflow/tool_handler.py are accurate and
verified still wired today: fact-forcing at tool_handler.py:3096 and config
protection at tool_handler.py:3061.

The pattern that already worked

#13590 (CLOSED) is the template. The repetition/stagnation guard had the identical
problem, and the fix was not to wire AgentLoop in — it was to extract the
decision logic into autobot_shared/repetition_guard.py as a dependency-free pure
function and call it from the live seam (tool_handler.py:3163). Its own module
docstring records the diagnosis: the machinery "existed in agent_loop/ and ran
nowhere". fact_forcing took the same route and is now a thin re-export of
autobot_shared.fact_forcing_guard.

Keep the decision logic pure and side-effect-free; let the graph node own the side
effect. That is what makes it both testable and wireable.

Scope

Extract PreActionVerifier's decision surface into autobot_shared/,
preserving the HARD_BLOCK verdict, and call it from the live approval gate.
Extract the belief-state update path the same way; wire it into the tool seam
where tool results are already collected.
Decide belief_state_enabled's default deliberately once it is reachable —
flipping it while unreachable changes nothing and would read as a fix.
Add a wiring assertion per guard: a test that fails if the live seam stops
calling it. Without this the next refactor silently un-wires it again, and
"no failures" reads as "working".

Related, same root cause, already filed: #13919 (read_spilled_output has no
production callers) and #13997 (tool-output spill wired only into AgentLoop).
Those two plus this one are the full inventory of agent_loop/ capability that
never reaches production.

Order

Wave B of #13587, after #13590 (CLOSED — it establishes the extraction
pattern). No dependency on #14027/#14028/#14029. Coordinate with #13919/#13997:
one extraction pass over agent_loop/ is cheaper than three.

Found during a comparative audit of an external agent-harness architecture, where
guards are pure decision classes plugged into graph nodes so there is only one
runtime and no second harness to fall out of date.

Full audit context: docs/research/agent-harness-guard-and-context-audit.md — landing in PR #14032.

The doc records what AutoBot already covers (with file:line evidence, so this issue is not re-audited later), where our design is deliberately stronger than the source's, and the do-not-re-file list — alongside all five gaps (#14027, #14028, #14029, #14030, #14031) and their wave order.

Correction — this is not a new discovery; #11221 already documented it

A code-reviewer pass caught a material omission in the body above.

The dormant loop is known, in-code, under a CLOSED issue

autobot-backend/agent_loop/__init__.py:11-15 already says:

NOT WIRED IN PRODUCTION (#11221): AgentLoop is never instantiated by any production
caller — the live tool seam is chat_workflow._dispatch_tool_call. Submodules here
(belief_state, extractors, guards) ARE reused by other subsystems, but the loop
orchestrator itself is dormant, so anything wired only into it is inert.

#11221 is CLOSED — "AgentLoop (agent_loop/) is never instantiated in production — decide
wire vs port vs document", resolved as document. The body above presents the dormancy as
freshly discovered while correcting #13587, and never cites #11221. That is exactly the kind
of omission that makes an issue look bigger than it is.

What is actually new — and it is narrower

pre_action_verifier_enabled defaults to True (agent_loop/types.py:266) inside that
dormant module. A disabled-by-default dormant component is honest; an enabled-by-default one
reads as an active guard while executing on no request — and is very likely why Umbrella: agent-seam governance is name-scoped, and the best guards are not on the production path #13587
miscounted it. belief_state_enabled defaults to False (:258), so it is off and
unreachable, which is the less misleading of the two.
Umbrella: agent-seam governance is name-scoped, and the best guards are not on the production path #13587's "Already covered — NOT filed" list miscounts both, citing agent_loop/loop.py
for the pre-action verifier and for durable run checkpointing. Documenting a dormant module
under AgentLoop (agent_loop/) is never instantiated in production — decide wire vs port vs document #11221 did not prevent a later audit from reading it as live coverage — which is the
real lesson, and the argument for extraction over documentation.

Grep wording corrected

grep -rn "from agent_loop.loop import\|AgentLoop(" returns two non-test hits besides the
tests: agent_loop/__init__.py:38 (package re-export) and :28 (docstring usage example).
Neither is an instantiation, so the conclusion holds — but "returns only test files" was wrong
as stated.

What this changes about the work

Nothing about the scope; something about the framing. #11221 chose document, and the
documentation demonstrably failed to stop a downstream audit from treating the module as live.
That is the case for choosing extract-and-wire this time rather than documenting again —
and for the wiring assertion in the acceptance criteria, which is the only form of
documentation that fails loudly.

Full corrected evidence: docs/research/agent-harness-guard-and-context-audit.md, Gap 5
(PR #14032).

docs(research): agent-harness guard, context-sizing and ingest-govern…

First slice landed: PR #14119 extracts and wires only PreActionVerifier — belief-state extraction and the belief_state_enabled default decision are deliberately left for a follow-up (belief-state's flag defaults False, so it's off and unreachable, a smaller/separable problem than the verifier's default-True-but-inert one).

Summary of #14119:

autobot_shared/pre_action_verifier_guard.py — the extracted decision surface (pure: threshold resolution, response parsing, determine_verdict/panel_decision — the HARD_BLOCK-preserving part) plus the PreActionVerifier class (the LLM call itself can't be made pure).
agent_loop/pre_action_verifier.py is now a thin re-export (the fact_forcing precedent).
chat_workflow/tool_handler.py gets a new _enforce_pre_action_verifier enforcer in the live dispatch chain, scoped to SENSITIVE_TOOLS, gated on pre_action_verifier_enabled (guard-profile resolved, defaults True).
Wiring assertion + mutation test included, per the feat(agent-loop): production seam has no repetition/stagnation guard — the guards exist in a dormant module, unported #13590 template.

Flagged in the PR body as a finding rather than a quiet scope change: execute_command is in SENSITIVE_TOOLS and is almost certainly the highest-volume sensitive tool in production, so this adds an LLM round trip in front of its existing approval flow for every shell command. That's the original AgentLoop scope exactly (never exercised before because AgentLoop never ran in production) — worth a latency/cost look post-merge, not something I narrowed unilaterally.

Leaving this issue open for the belief-state slice.

fix(agent): prove the verifier wiring behaviourally and survive a ctx…

fix(agent): put the F822 suppression where flake8 reports it (#14031)

fix(agent): route the verifier's role literals through the SSOT const…

fix(agent): propagate a cancelled panel task instead of unpacking it (#…

fix(agent): extract and wire the pre-action verifier at the live tool…

Half of this has been fixed since filing — one of the two components named is now live.

PreActionVerifier — no longer dead. It has a real non-test caller:

autobot-backend/chat_workflow/tool_dispatch_guards.py:193
    await PreActionVerifier().verify(...)

imported by chat_workflow/tool_handler.py:51, plus the extracted autobot_shared/pre_action_verifier_guard.py:387. This landed via the #13590 extraction template.

BeliefStateUpdater and AgentLoop — still dead. Non-test references to BeliefState are agent_loop/loop.py:28,186 (its own definition and use) and benchmarks/benchmark_belief_state.py:78,384, which is a benchmark script, not production. For AgentLoop:

$ git grep -n -E "from agent_loop.loop import|AgentLoop\(" origin/Dev_new_gui | grep -v test
0 non-test callers

The single hit in agent_loop/__init__.py:28 is inside the module docstring's usage example, not code.

So the remaining scope is BeliefStateUpdater + AgentLoop, and the title's "pre-action verifier and belief state run nowhere" is now half true. Per the never-delete rule these are unfinished work to wire in, not code to remove — and PreActionVerifier is the worked example of how.

Verified against merged main. One of four delivered; the other three all hang off the same missing piece.

AC1 — delivered, ticked

28e9df7c8 (PR #14119) extracted the verifier and wired it at the live seam:


`autobot_shared/pre_action_verifier_guard.py`	564 lines — the decision surface, out of `agent_loop/`
`autobot-backend/chat_workflow/tool_handler.py`	+99 lines — the seam now calls it
`autobot-backend/agent_loop/pre_action_verifier.py`	−465 lines — the dead copy shrank to a shim

So the verifier no longer "runs nowhere": the tool path invokes it.

AC4 — half delivered

autobot-backend/tests/unit/chat_workflow/test_tool_handler_pre_action_verifier.py carries test_the_call_site_still_names_the_enforcer, which is exactly the wiring assertion this criterion asks for — it fails if the seam stops calling the guard.

But the criterion says per guard, and there is only one guard wired. The belief-state half has no seam to assert against, so AC4 cannot complete before AC2.

AC2 and AC3 — not delivered

belief_state appears in no autobot_shared/ or chat_workflow/ file on main. Every reference is still inside autobot-backend/agent_loop/:

agent_loop/__init__.py   agent_loop/loop.py   agent_loop/types.py
agent_loop/text_utils.py agent_loop/tests/test_belief_cache.py
benchmarks/benchmark_belief_state.py

AgentLoop still has no production caller, so belief state is in exactly the position this issue was filed about — the verifier was lifted out of it and the belief-state path was not.

AC3 (decide belief_state_enabled's default deliberately once it is reachable) is explicitly conditional on that reachability, so it cannot be judged, let alone met, until AC2 lands.

One thing worth a look

28e9df7c8 also added a tracked entry named venv (venv | 1 + in its diffstat). That looks like an accidental commit of a virtualenv symlink rather than intended content. Not this issue's subject and I have not touched it — flagging it because a tracked venv entry tends to cause confusing failures later.

Remaining work is one extraction (AC2), which then unblocks AC3 and completes AC4. Not closing.

Correction to my previous comment's last section: the venv entry is already fixed and needs no look.

I read it out of 28e9df7c8's diffstat and flagged it without checking the present state. It is not on main: d74d66875 — "fix(repo): untrack the venv symlink that destroys the interpreter on checkout (#14133) (#14138)" — removed it, and it had its own issue.

So that paragraph described a commit correctly and the current tree wrongly. The two are different claims and I made the second while only checking the first. Everything above it stands: AC1 delivered and ticked, AC4 half delivered, AC2 and AC3 blocked on the belief-state extraction.

Re-verified against current `main` — the defect holds, and re-checking it has a trap in it

Triaged under the branch freeze.

The claim still holds

AgentLoop has no production instantiation on main. PreActionVerifier and
BeliefStateUpdater remain reachable only through agent_loop/loop.py, which nothing constructs:

PreActionVerifier — loop.py:28 (import), loop.py:191-192 (construction)
BeliefStateUpdater — loop.py:27, loop.py:186; the only other non-test reference is
benchmarks/benchmark_belief_state.py:78,384, a benchmark, not a caller
Defaults unchanged: types.py:266 pre_action_verifier_enabled: bool = True,
types.py:258 belief_state_enabled: bool = False

So the verifier still reads as an active guard — enabled by default — while executing on no
request at all. That is the sharpest part of this issue and it is still true.

The trap, worth recording before someone re-verifies

The issue's evidence section says the grep returns "only test files". Run today, it also returns:

autobot-backend/agent_loop/__init__.py:28:    loop = AgentLoop(

That is inside the module docstring — the package's own usage example, not code. A reader
checking whether this issue is stale sees a non-test hit in a production file and can reasonably
conclude a caller now exists and the issue is fixed. It is not; the only other reference at
__init__.py:38 is the re-export.

This is the same failure this issue is about, one level up: something that reads as reachable while
reaching nothing. Anyone re-verifying should exclude docstrings, or check what the line is inside
before counting it.

Branch verdict: do not revive

issue-14031 is 4,286 files divergent from main, 6 weeks old, with two unlanded paths:

autobot-backend/chat_workflow/tool_handler.py — its copy predates security(agents): an agent refused an action can get another agent to do it — no per-peer identity #16950 and feat(agents): deliver peer messages at the recipient's next turn, not mid-task #16948, so taking
it would drop both (guards(git): reviving a stale branch can silently reverse a later fix — 75 of 87 revivable branches carry one #17178).
venv — a committed symlink (mode 120000) pointing at an absolute path inside a
developer's home directory. It is absent from main, correctly, and must stay that way.

The second one is worth naming: it is invisible in a file-count summary, survives any "only two
files changed" reading, and would plant an absolute machine-specific path in the repository. It is
also a second reason this branch is not a starting point for anything.

What the fix still needs

Per the never-delete rule, both components get wired into the live path —
chat_workflow/graph.py and chat_workflow/tool_handler.py — rather than removed. Until then the
honest interim state would be for pre_action_verifier_enabled to stop defaulting to True, since
a default that advertises an inactive guard is the misleading half of the defect.

Issue stays open. Nothing here is recoverable from the branch.

Correction: the verifier half of this issue is resolved

Re-checked on main during a comparative research audit
(docs/research/confidence-gated-decision-loop-harness.md).

PreActionVerifier now has a production caller.
chat_workflow/tool_dispatch_guards.py:154 enforce_pre_action_verifier, whose own docstring
records it as "its first production caller", wired into the dispatch chain at
chat_workflow/tool_handler.py:3264-3335. Scope matches the original — only tools matching the
canonical SENSITIVE_TOOLS set — and on a soft block it threads verifier_rationale,
verifier_refutation_probability and verifier_degradation into the pending_approval payload,
so the rationale reaches the human.

Still true: AgentLoop itself has no production caller. grep -rn "AgentLoop(" outside
tests returns one hit, the docstring usage example at agent_loop/__init__.py:28.

Not re-checked: the belief-state half of this issue.

Adjacent defect found in the same config object and filed separately as #17469 — the three
timeout budgets in agent_loop/types.py:189-192 are read by nothing, so wiring this loop would
not bring its budgets with it.

docs(research): carry the confidence-gated harness handoff too (#17471)

docs(research): land two comparative audits whose 12 issues are alrea…

v0.9.1 triage (classification only — not a fix proposal)

Pre-filter — PRs: fix(agent): extract and wire the pre-action verifier at the live tool seam (#14031) #14119 MERGED 2026-08-12 "extract and wire the pre-action verifier at the live tool seam (fix(agent): pre-action verifier and belief state run nowhere — AgentLoop has no production caller #14031)" — targets this (AC1 slice). docs(research): agent-harness guard, context-sizing and ingest-governance audit (#14027) #14032 MERGED "docs(research): agent-harness guard… audit (security(commands): dangerous-command patterns match un-normalized input — homoglyph and ANSI-wrapped commands bypass the gate #14027)" — mentions only.
Pre-filter — commits since filing (2026-08-11): --grep '#14031': 28e9df7c8 (fix(agent): extract and wire the pre-action verifier at the live tool seam (#14031) #14119), 47238c619 (docs(research): agent-harness guard, context-sizing and ingest-governance audit (#14027) #14032 doc), 94c427df3 (docs, mention). Path log on agent_loop/{loop,types,pre_action_verifier}.py, autobot_shared/pre_action_verifier_guard.py, chat_workflow/tool_dispatch_guards.py: 28e9df7c8, fb7f85f71 guards extracted from tool_handler (bug(ci): Dev_new_gui is red — tool_handler.py grew past its #14236 ceiling (4081 vs 4063), failing code-quality on every open PR #14495), c041c120d env casts, 3d9f3ba9a (security(agents): an agent refused an action can get another agent to do it — no per-peer identity #16950), de264fd15 (feat(agents): deliver peer messages at the recipient's next turn, not mid-task #16948/9), e8c7bf9b5 degraded verifier replies (bug(llm): structured_output=True is honoured by exactly one provider — every hosted provider silently drops it #17305/6). None touches belief state.
Files: autobot-backend/agent_loop/types.py (grep :258,:266), agent_loop/loop.py (grep belief refs), agent_loop/__init__.py (opened :20-40), agent_loop/belief_state.py (exists, not opened), chat_workflow/tool_dispatch_guards.py (grep :154-197), chat_workflow/tool_handler.py (grep :56,:3226-3236,:3266,:3319), tests/unit/chat_workflow/test_tool_handler_pre_action_verifier.py (test names), benchmarks/benchmark_belief_state.py (grep).
Premise check (code read on main 80b6200):
- "Pre-action verifier runs nowhere" — STALE (AC1 met, ticked): real call tool_dispatch_guards.py:197 await PreActionVerifier().verify(...) inside enforce_pre_action_verifier :154, reached from tool_handler.py:3319 inside _dispatch_tool_call (def :3266) via _enforce_pre_action_verifier :3226-3236.
- "AgentLoop has no production caller" — TRUE: non-test hits are agent_loop/__init__.py:28 (docstring usage example) and :38 (re-export) only.
- Belief state reachable only from AgentLoop — TRUE: loop.py:27 import, :186 construction, :644,:699,:1350,:1385 use; only other non-test reference is benchmarks/benchmark_belief_state.py:78,384. No belief_state in chat_workflow/ or autobot_shared/.
- Defaults — TRUE, unchanged: types.py:258 belief_state_enabled = False, :266 pre_action_verifier_enabled = True (the latter now governs a live path, so the "reads active while inert" part applies only to the AgentLoop copy).
- AC4 (wiring assertion per guard): verifier half present — behavioural test_the_live_dispatch_seam_reaches_and_awaits_the_verifier (:80) plus rename guard (:63); belief-state half impossible until AC2.
- Comment 5751505936 (2026-09-20) says PreActionVerifier is still reachable only via loop.py — contradicted by the lines above; later comment 5828871088 (2026-09-25) corrects it.
Remaining scope per code: AC2 (belief-state extraction + seam wiring), AC3 (default decision, conditional on AC2), AC4 belief half.
Blocked by: nothing hard. Issue asks to coordinate with fix(agent): wire read_spilled_output into the production tool seam — execute_tool has no production callers #13919 and fix(agent): the tool-output spill is wired only into AgentLoop, which is never instantiated in production #13997 (one extraction pass over agent_loop/); adjacent bug(agent-loop): three timeout budgets declared in AgentLoopConfig and read by nothing #17469 (unused timeout budgets in types.py:189-192, per comment). AC3 is a deliberate owner decision gated on AC2.
Read vs inferred: Read: body + 10 comments; lines cited above. Inferred: that no belief-state seam exists under another name (searched BeliefState|belief_state only, not the concept "assertions tracker").
Undetermined: state of fix(agent): wire read_spilled_output into the production tool seam — execute_tool has no production callers #13919/fix(agent): the tool-output spill is wired only into AgentLoop, which is never instantiated in production #13997 (not opened); whether the belief-state concept exists in the live graph under different naming.
Collisions within v0.9.1 (same file; one PR/one agent under the same-file rule): autobot-backend/chat_workflow/tool_dispatch_guards.py with feat(agents): deliver peer messages at the recipient's next turn, not mid-task #16948, AI_STACK peer-message drain only wired for the 'chat' role; 'rag'/'system_commands' silently lose messages #16997

Generated by Claude Code

Pre-filter + per-AC pass — partial, and the gap is extraction rather than wiring. Not closed.

Read-only at origin/main ab05ba3aef. One commit names this issue: 28e9df7c84 "fix(agent): extract and wire the pre-action verifier at the live tool seam (#14031)" — which is AC1, already ticked.

The belief-state half (AC2) is wired but not extracted. belief_state_enabled is consulted in the live loop at agent_loop/loop.py:644, :699 and :1350, so the path is reachable — but the criterion asks to "extract the belief-state update path the same way", i.e. into autobot_shared/, as the verifier was. A search for belief_state/BeliefState under autobot_shared/ finds nothing. So the asymmetry this issue was filed about is half closed: the verifier is shared and wired, the belief state is wired in place.

That distinction matters for AC4 (a wiring assertion per guard: a test that fails if the live seam stops). A guard over a shared module can assert the seam from outside; a guard over a path that only exists inside loop.py has to assert against the loop, which is the shape that rots. Extraction is the precondition for the guard, not a tidiness step — which is presumably why the criterion says the same way.

AC3 (decide belief_state_enabled's default deliberately once it is reachable) is now answerable, since the three call sites above make it reachable. It is a decision, not work: the default is currently whatever the config says, and the criterion asks for that to be a recorded choice rather than an inherited one.

Verdict: 1 of 4, with AC2 as the blocker for AC4 and AC3 available to decide now. Nothing here needs a host.

B7 spec — #17725 delivers most of this issue's open criteria. The gap is narrower than the batch plan assumed.

Full spec at ~/b7-spec-2026-09-29.md. Read from origin/issue-16818-lease-wiring (PR #17725, open) against origin/main.

I expected to specify the release path. The branch already contains it — its +28 to stalled_run_sweep.py is:

+sweep now releases what the run held: a stalled run hands its workspace back on the
+same pass that closes it out.
+from llc.services.workspace_lease import release_for_run
+STALL_RELEASE_REASON = "run stalled: workspace released by the stalled-run sweep (#16817/#16818)"
+    released += len(await release_for_run(session, run.id, STALL_RELEASE_REASON))

AC	After #17725 merges
AC2 releases what the run held	lease: delivered
AC5 two sweepers must not double-release	delivered, with the reasoning in the diff — "Inside the same transaction as the status write, and under the same SKIP LOCKED selection: a run that this worker did not claim is not this worker's to release"
AC6 release half of the test	`stalled_run_sweep_test.py` +56
AC7 reports what it released	`released` counter + `logger.info`

The gap: AC2 names three things and the branch releases one

"claims, assignments, workspace leases" — no claim or assign appears anywhere in those additions.

Change: in the same transaction and under the same SKIP LOCKED selection, release the run's claims and assignments alongside its lease, each with a reason distinguishable from an expiry the way STALL_RELEASE_REASON already is.
Assertion: a swept run afterwards holds no claim, no assignment and no live lease, and the log names all three counts.
Mutation: release the lease only → the claims/assignments assertion reddens while the lease assertion stays green. Without that divergence, "releases what the run held" is satisfied by releasing one of three things.

One boundary #17725 records, which this must not cross: "The release frees the slot, not the directory. Whether the directory may be removed is a separate question with its own evidence, and it is asked in llc.services.workspace_disposal." Directory disposal is out of scope here.

Navigation Menu

Repository navigation

fix(agent): pre-action verifier and belief state run nowhere — AgentLoop has no production caller #14031

Problem

Evidence

Correction to this umbrella's own premise

The pattern that already worked

Scope

Order

Correction — this is not a new discovery; #11221 already documented it

The dormant loop is known, in-code, under a CLOSED issue

What is actually new — and it is narrower

Grep wording corrected

What this changes about the work

3 remaining items

AC1 — delivered, ticked

AC4 — half delivered

AC2 and AC3 — not delivered

One thing worth a look

Re-verified against current `main` — the defect holds, and re-checking it has a trap in it

The claim still holds

The trap, worth recording before someone re-verifies

Branch verdict: do not revive

What the fix still needs

Correction: the verifier half of this issue is resolved

v0.9.1 triage (classification only — not a fix proposal)

Pre-filter + per-AC pass — partial, and the gap is extraction rather than wiring. Not closed.

B7 spec — #17725 delivers most of this issue's open criteria. The gap is narrower than the batch plan assumed.

The gap: AC2 names three things and the branch releases one

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

fix(agent): pre-action verifier and belief state run nowhere — AgentLoop has no production caller #14031

Description

Problem

Evidence

Correction to this umbrella's own premise

The pattern that already worked

Scope

Order

Activity

mrveiss commented on Aug 11, 2026

mrveiss commented on Aug 11, 2026

Correction — this is not a new discovery; #11221 already documented it

The dormant loop is known, in-code, under a CLOSED issue

What is actually new — and it is narrower

Grep wording corrected

What this changes about the work

mrveiss commented on Aug 11, 2026

3 remaining items

mrveiss commented on Sep 1, 2026

mrveiss commented on Sep 17, 2026

AC1 — delivered, ticked

AC4 — half delivered

AC2 and AC3 — not delivered

One thing worth a look

mrveiss commented on Sep 17, 2026

mrveiss commented on Sep 20, 2026

Re-verified against current main — the defect holds, and re-checking it has a trap in it

The claim still holds

The trap, worth recording before someone re-verifies

Branch verdict: do not revive

What the fix still needs

mrveiss commented on Sep 25, 2026

Correction: the verifier half of this issue is resolved

mrveiss commented on Sep 28, 2026

v0.9.1 triage (classification only — not a fix proposal)

mrveiss commented on Sep 28, 2026

Pre-filter + per-AC pass — partial, and the gap is extraction rather than wiring. Not closed.

mrveiss commented on Sep 29, 2026

B7 spec — #17725 delivers most of this issue's open criteria. The gap is narrower than the batch plan assumed.

The gap: AC2 names three things and the branch releases one

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions

Re-verified against current `main` — the defect holds, and re-checking it has a trap in it