Repository navigation
fix(aidd-context): source-backed tool references, with dates and unverified cells #618
Description
Activity
Corrections mesurées, 2026-08-13/14
Campagne de sondes sur les cinq outils, sessions réelles et documentations officielles. Six corrections à porter dans
tool-paths.md, toutes vérifiables.Sujet État du fichier Mesuré Codex SessionEndabsent du tableau des moments existe. 1 s par défaut, 3 s au maximum, ne se déclenche pas pour les sous-agents Moments Codex huit listés trois manquants : PermissionRequest,PostCompact,SubagentStartVerrous de hook rien du tout voir ci-dessous Cursor en ligne de commande non traité cursor-agentlit bien.cursor/hooks.json— mesuré, la doc ne décrit que des moments d'éditeurIdentifiants Cursor un seul décrit trois dans le payload : session_id,conversation_id,generation_id, égaux sur une session à un tourOpenCode « hooks impossibles » exact pour les hooks, mais il exporte en OpenTelemetry derrière experimental.openTelemetrydansopencode.jsonLe manque le plus important : les verrous
Aucune sonde n'a fonctionné du premier coup, et jamais pour une raison différente. Écrire le fichier de hook ne suffit pas.
Outil Verrou Levée Codex drapeau de fonctionnalité et confiance persistée --enable hooks+--dangerously-bypass-hook-trustCopilot confiance du dossier pour .github/hooks/*.jsonpérimètre utilisateur $COPILOT_HOME/hooks/, non soumis à la confianceCursor confiance de l'espace de travail --trustou approbation interactiveClaude Code aucun — Un hook posé sans lever le verrou est installé, silencieux, et ne lève aucune erreur. C'est le pire état pour une couche de mesure : la configuration paraît complète et la donnée n'existe pas. Ce tableau manque au fichier et conditionne #617.
Sur OpenCode, correction d'une correction
La ligne « OpenCode n'a aucune télémétrie » était fausse, et je l'ai écrite deux fois avant de la mesurer. La bascule est une clé de configuration, pas une variable d'environnement. Activée, ses spans
ai.streamTextportentgen_ai.usage.input_tokens,gen_ai.usage.output_tokensetsession.idsur le même span. Il honore aussiOTEL_RESOURCE_ATTRIBUTES. Réserve : 495 spans pour une session triviale, tout étant instrumenté jusqu'aux lectures de fichier.Détail complet et sources dans
aidd_docs/brainstorm/2026_08_13-telemetry-layer.md.- added a parent issue
on Aug 14, 2026 Out of the week's scope, 2026-08-16
Milestone 14 is scoped to what produces a trustworthy figure in five days: #620 the journal, #646 the switch, #647 the sink, #663 the step boundaries, #629 the figure, #617 the proof it is honest. Plus #658, which ships with them because the published documentation currently promises the opposite of what we are about to release.
This issue is not in that set. It keeps its value and its content; it simply is not what the week is measured on. It returns to the milestone board once the first figure exists.
No milestone is assigned rather than a later one, because where it lands depends on what the first week actually teaches.
I suggest keeping this open as a bug, with a narrower first pass.
tool-paths.mdstill states that Codex has noSessionEnd, and it still has neither verification marks nor source links. Current Codex documentation confirmsSessionEndfor the main thread, not subagents. The 1-second default and 3-second maximum found in the earlier note apply toInterrupt, so they should not be copied toSessionEndwithout a direct source.First deliverable: correct
tool-paths.mdand add a compact evidence record for each factual claim: source URL, checked date, and verified or unverified status. Validate that shape there before extending it to the other mappings.I would keep this outside a milestone as decided, but set its Roadmap priority to High: this reference drives generated hook configurations.
- added 2 commits that reference this issue
on Sep 23, 2026
Problem
plugins/aidd-context/skills/08-hook-generate/references/tool-paths.mdstated that Codex CLI has no session-end moment. Codex documentsSessionEndexplicitly. The claim was wrong, and it propagated: a downstream design decision was taken on it before anyone re-read the source.The failure is structural, not a typo. Not one cell in that table carried a URL or a date. Nothing distinguished "read in the docs" from "assumed in 2026-06 and never re-checked", so the only way to audit one cell was to re-verify all of them. A reference that cannot be audited cell by cell decays silently, and agents quote it with full confidence.
These references are load-bearing: a skill reads them instead of the vendor docs, so an error there becomes an error in every artefact the skill generates.
Scope
Applies to reference files that make per-tool factual claims, starting with
08-hook-generate/references/tool-paths.md.[v]= read in the official docs on the stated date.[?]= not re-read, unverified. An unverified cell stays unverified; it is never filled with a plausible value.developers.openai.com/codex/hooksanddocs.github.com/.../hooks-configurationboth moved).[v]becomes stale and must be re-read.aidd-contextonce the shape is agreed (11-explore/references/ai-mapping.md,03-context-generate/references/ai-mapping.md).A first pass over
tool-paths.mdis written and ready to land with this issue: marks, sources, verification date, plus the corrections below.Corrections found while verifying (5 tools, official docs, 2026-08-11)
SessionEnd. Previously recorded as absent. Timeout is 1 s by default, 3 s max, and it does not run for subagents.agentStopbecomesStop,userPromptSubmittedbecomesUserPromptSubmit. The file described it as accepting both names.preToolUsehandler that exits non-zero on Copilot denies the tool call. An observer hook that throws stops the agent. Not documented anywhere in the file before.session_id,conversation_id,sessionId) and none of the five documents whether the value survives a resume, clear, compaction or fork.Acceptance criteria
tool-paths.mdcarries[v]or[?].Sourcessection lists each page read, with the date of the pass.[?]cell is never silently promoted to[v]without a re-read.Prior art in this repo
plugins/aidd-context/skills/08-hook-generate/references/tool-paths.md— the file this issue fixes.plugins/aidd-context/skills/08-hook-generate/references/hook-authoring.md— R1-R7, which delegate every per-tool fact totool-paths.md.AGENTS.md— "Don't guess APIs, signatures, flags, or behavior" and "Don't assume your knowledge is current" already carry the intent. This issue does not add a rule; it makes the existing ones auditable in the file that holds the claims.Relations
Verified against: Claude Code hooks and monitoring-usage, Codex hooks and config-advanced, Cursor hooks, Copilot hooks-reference, OpenCode plugins and config schema.