Repository navigation
docs(memory): a canary is a regression test, not a plausibility check - #373
Merged
Merged
Conversation
CLOUD-315 The surveyed sources are gated, and the survey's reasoning has a home
WhyA survey of the commercial static-analysis and agent-hook field produced three follow-on issues and a set of judgements that belong to no issue: why four of five candidates were rejected, and what to do differently next time a tool is evaluated. Two things follow, and skipping either makes the survey a half-change. The reasoning needs its destination. Per the sorting rule it is neither an issue nor a rule that binds every turn, so it goes to the memory that owns surveys — which is also the only file the attribution gate exempts, and therefore the only place a surveyed source may be named. The gate needs the new names. Mechanism
Ready
Done
|
wenzowski
marked this pull request as ready for review
August 12, 2026 19:19
wenzowski
marked this pull request as draft
August 12, 2026 19:20
wenzowski
force-pushed
the
claude/semgrep-guardian-research-t6va58
branch
from
August 12, 2026 20:09
2cda6b4 to
2df4050
Compare
A survey that classifies a corpus needs cases whose answer is known because getting them wrong is a mistake already made, and the run must fail when one misclassifies. Round one of this survey matched policy files by filename and was wrong in both directions; round two was built to make that unrepresentable and reproduced the same blind spot in a new costume, fetching only paths guessed in advance. The canaries are what caught it. The transferable half is the sentence about instruments: absence of evidence is a claim about the instrument before it is a claim about the world. The corollary for any classifier is that a name may decide what gets read, and only the text decides what a thing is. Refs: CLOUD-315
wenzowski
marked this pull request as ready for review
August 12, 2026 21:41
wenzowski
force-pushed
the
claude/semgrep-guardian-research-t6va58
branch
from
August 12, 2026 21:41
2df4050 to
bd1c7c5
Compare
The command shipped with the framing the memory beside it had already dropped: "the honest signal that you are past it is a rising re-verify rate", plus a pointer to `mem:workflow/agent-fanout` for "the current cap", which that memory no longer states in those terms. Two authorities disagreeing about the same fact, in the file whose §1 was "one authority each, no duplication" — and the command is the reachable one, so a dispatcher invoking `/plan-fleet` got the superseded model while the memory it points at carried the current one. Replaced with the cost table the memory already holds: re-verifying is free because `land` laps without a model turn, and what actually costs is a rebase conflict, CI minutes on a voided run, and tokens. Objective restated as pace of landed work per token, with the saturated-queue target and the CSMA/CD framing, and the CI-minutes row now names CLOUD-369 as unbuilt rather than implying a control exists. Refs: CLOUD-367
…nheriting a spent window Two defects in `plan-hold`, both measured 2026-08-12 in one session. THE ANSWER WAS INVISIBLE. `plan-hold-guard` gates `PreToolUse` on `ExitPlanMode|AskUserQuestion`; `plan-hold-release` listens on `UserPromptSubmit`. Those are different event classes, and a human answering either of those two tools produces a TOOL RESULT, not a prompt. So for the exact case the mechanism exists to serve, the release could never fire: a hold armed, an `AskUserQuestion` answered, the hold still live afterwards, ended by removing its sentinel by hand. CLOUD-451's third acceptance bullet, unmet since it landed. `plan-hold-release-tool` is the missing half, on `PostToolUse` over the same two tools the guard already names. No classifier: `plan-hold-release` must decide whether a prompt came from a person because `UserPromptSubmit` carries machine turns too, while here the provenance is structural — the event fires only after a tool whose whole purpose is to ask a human. The `tool_name` check is belt to the matcher's braces, because a matcher widened later would otherwise turn every tool call into a release. CLOUD-435's cost argument does not reach it for the reason it does not reach `plan-hold-guard`: these two tools fire at most once per turn, and it is invoked by path, so no task-runner startup is paid. THE WINDOW ERODED SILENTLY. A second launch answered "already held" and exited, which keeps one sleeper and hands the new handoff whatever is left of the old one's cap. Some answers reach neither release path — a reply typed mid-turn arrives `<`-wrapped and reads as a machine turn — so that handoff's hold never releases, the next inherits a partly spent window, and eventually one is guarded by a hold about to cap and the container is reclaimed while somebody is reading. Nothing blocks, so nothing announces it. A second launch now releases the incumbent and arms fresh: at most one sleeper still, and every handoff gets the whole cap. Released, never signalled — a killed hold wakes nothing. Deliberately NOT touching `plan-hold-release-check`'s `<`-prefix classifier. The mid-turn false negative is real, but widening a literal without measuring it over real transcripts is the mistake CLOUD-252 and CLOUD-323 exist to prevent. The erosion fix removes its compounding cost; the false negative itself stays recorded on CLOUD-485. Refs: CLOUD-485
wenzowski
marked this pull request as draft
August 12, 2026 21:50
Every bound in this memory was a bound on landing, and its cap of 2 is derived from land contention. An agent that researches, reviews, or drafts never lands, so that derivation says nothing about it — and the silence read as absence of a constraint rather than absence of a derivation. Eight drafting agents were launched into it. The binding constraint for that shape is per-agent fixed cost, which is paid once per agent and multiplies with N. The measured floor now has a home: 63,848 tokens for one agent to fetch a single issue and run one lint, for near-zero work. Each prompt named eight artifacts as required reading, so that floor was paid eight times over before anything happened. Three rules follow — digest rather than cite, checkpoint at each unit's own gate because scratch dies with the container, and pilot one unit through its durable write before widening — and the section says outright that none of them is gated, because the predicate they need measures a session's real spend and that is out of tree (CLOUD-95). Assumed enforcement is worse than stated absence. Also corrects the implementer contract in the same file: it still listed a hand `gh pr ready` before `land`, which is the defect CLOUD-247 named, and `land` has been the only readier since. Refs: CLOUD-289
|
This was referenced Aug 13, 2026
Merged
This was referenced Aug 26, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What changed
One paragraph added to
mem:prior-art-and-issue-hygiene, recording what the field survey taught about instruments.Why it mattered here
The survey classified 20,139 repositories. Round one matched policy files by filename and was wrong in both directions — a
deny.tomlthat belonged to cargo-deny, aconftest.pythat belonged to pytest. Round two was built specifically to make that unrepresentable, and reproduced the same blind spot in a new costume: it fetched only the paths guessed in advance, so root-level policy files nobody thought to name stayed invisible.What caught it both times was a small set of cases whose correct answer was already known by hand, asserted as a hard failure rather than reviewed by eye.
What a reviewer should look at
The transferable claim, which is the only reason this is durable rather than a session note: absence of evidence is a claim about the instrument before it is a claim about the world. The corollary for any classifier is that a name may decide what gets read, and only the text decides what a thing is.
Refs: CLOUD-315