intent-check: request standard effort/profile instead of frontier/maximum - #114
intent-check: request standard effort/profile instead of frontier/maximum#114sumitake wants to merge 1 commit into
Conversation
…imum The Step-0 intent check is a bounded two-document interpretation comparison, not architecture-grade deliberation, yet it hardcoded quality_profile:frontier + effort_class:maximum in its coordinator request -- making it the single most expensive high-frequency managed call in the system with no quality justification. For the context.documents.intent route the quality_profile knob is functionally inert (identical candidate lists across economical/standard/ frontier), so this is a pure cost reduction with no routing change. Matches the existing teamwork/orchestrate standard/standard default pattern. Edited at source (skill-specs/intent-check.md + skill-build-config.json's intent_check_defaults_block) and regenerated SKILL.md via build_skills.py. No version bump -- the in-flight v6.0.6 release owns versioning (operator direction); this fragment compiles into that release. Companion: sumitake/agent-collab-workspace#2770 downgrades the route-level effort_class floor. Workspace merges FIRST -- a request below the floor is rejected effort_ineligible, so merging this first would break intent-check.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: de21ea0b80
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
|
||
| ```json coordinator-request | ||
| {"request_id":"intent-check-1","logical_action":"context.documents.intent","quality_profile":"frontier","effort_class":"maximum","target_agent":null,"timeout_ms":120000,"prompt":"Compare the operator request with the primary interpretation. Identify only material omissions, added scope, or ambiguity; do not design or execute the work.","documents":[{"label":"operator_request","content":"<verbatim operator request>"},{"label":"primary_interpretation","content":"<plain-language interpretation>"}]} | ||
| {"request_id":"intent-check-1","logical_action":"context.documents.intent","quality_profile":"standard","effort_class":"standard","target_agent":null,"timeout_ms":120000,"prompt":"Compare the operator request with the primary interpretation. Identify only material omissions, added scope, or ambiguity; do not design or execute the work.","documents":[{"label":"operator_request","content":"<verbatim operator request>"},{"label":"primary_interpretation","content":"<plain-language interpretation>"}]} |
There was a problem hiding this comment.
With the currently packaged 6.0.5/runtime-4.0.4 bundle, this request is below the existing context.documents.intent maximum-effort floor and is therefore rejected as effort_ineligible, so every invocation of intent-check from this tree fails before inference. Merging the companion workspace policy first does not update the standalone runtime bytes selected beneath this plugin root; this change must remain paired with importing the rebuilt, signed bundle that contains the lowered floor (or retain maximum until that bundle is packaged).
AGENTS.md reference: AGENTS.md:L13-L16
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Confirmed correct — verified the plugin bundle is a signed native runtime (runtime-manifest.json lists compiled .so artifacts with sha256 + Developer-ID signatures), so the routing floor is baked into signed bytes that the workspace policy change does not update. Merging this skill change against the current 6.0.5/4.0.4 bundle would reject the standard request as effort_ineligible before inference.
Action: holding this PR (back to draft). Per your option (a), the skill-request lowering must ship paired with the rebuilt/re-signed runtime bundle that carries the lowered intent floor — i.e. it belongs in the in-flight v6.0.6 release, not a standalone merge. The workspace source that feeds that rebuild is landing first (agent-collab-workspace#2771, the verify-intent carve-out). Good catch — this prevented a break-before-inference regression.
🤖 Addressed by Claude Code
|
Workspace companion updated: the original #2770 (option 1) was closed by operator decision; the companion is now sumitake/agent-collab-workspace#2771 (option 2 — the surgical |
|
Closing per operator correction: this plugin repo is built/imported from agent-collab-workspace — nothing is consumed from standalone merges to this repo's main, and release imports overwrite content wholesale. A standalone merge here would also break intent-check against the current 6.0.5 runtime (Codex P1: request below the baked-in floor -> effort_ineligible). The change survives as companion source branch |
1 participant
Summary
Removes the intent-check skill's hardcoded
quality_profile:frontier/effort_class:maximumcoordinator-request pin; it now requestsstandard/standard. The Step-0 intent check is a bounded two-document interpretation comparison, not architecture-grade deliberation, so the top-tier pin made it the single most expensive high-frequency managed call in the system with no quality justification.context.documents.intentroute thequality_profileknob is functionally inert (identical candidate lists across economical/standard/frontier in provider-routing.yaml), so this is a pure cost reduction with no routing change — same reviewer pool (gemini/grok), lower effort.teamwork/orchestratestandard/standarddefault pattern already inskill-build-config.json.skill-specs/intent-check.md+skill-build-config.json'sintent_check_defaults_block) and regeneratedSKILL.mdviabuild_skills.py(never hand-edit the generated file). Updated the one test that pinned the old value (test_unified_skill_runtime_contract.py) — the assertion still fully checks the field and the complete key-set, just against the new value.changelog.d/fragment that compiles into that release.Cross-repo coordination
Companion: sumitake/agent-collab-workspace#2770 downgrades the route-level
effort_classfloor inprovider-routing.yaml. Merge order is load-bearing: the workspace PR MUST merge first. The routeeffort_classis a floor (provider_action_eval.py:288) — a request below it is rejectedeffort_ineligible. If this plugin PR merged first, the now-standardrequest would fall below the still-maximumfloor and intent-check would break outright.Compliance trace
author: claude
standing_directives: directive #4 (project->plugin sync), directive #6 (cross-repo coordination), directive #7 (model+effort pairing), directive #8a (delegation economy)
tier: 3
cross_check: PENDING -- managed reviewer runtime unavailable this session (installed plugin cache pinned to broken v6.0.5; see agent-collab-workspace .learnings/entries/ERR-20260818-024526-claude-9e0a.md; requires operator session restart before cross-check / peer review can run)
post_condition: source content change only; no version bump (in-flight v6.0.6 release owns versioning per operator); local test gate (tests/ + scripts/) green pre-push; self-merge withheld pending cross-check + distinct-family peer review
mcp_coverage_gap: none
operator_reserved: false
Test plan
build_skills.py --check— generated SKILL.md matches source specbuild-changelog.py --dry-run— fragment compilescheck-public-export-safety.py --active-tree— SAFEpython3 -m unittest discover -s tests— 175 tests OK (updated the one stale assertion)