peer-review-agency-scope

Metadata

Description

## Description
Cross-model peer review of `impl-agency-scope-rules`. Verify scope-aware composition and composition-rules.csv overlay.

## What to check
- `scope` field on primitives respects agency's exact value set: task | meta:assigner | meta:evaluator | meta:evolver | meta:agent_creator (no synonyms, no typos)
- composition-rules.csv columns match agency exactly: agent_type, rule, max_role_components, max_desired_outcomes, max_trade_off_configs, all_projects, project_ids
- Watched-file semantics: edits are picked up without a daemon restart (live smoke output should show this)
- Back-compat: existing primitives with no `scope` field default to `task` and still compose correctly
- Composition caps actually constrain assignment — confirm via test that an over-cap composition gets pruned
- File scope respected: no schema field changes (those are in impl-agency-schema-fields), no hash changes

## Deliverable
`wg log` concur / concern verdict.

## Validation
- [ ] Read the composer change in src/agency/prompt.rs
- [ ] Read composition-rules.csv parser code
- [ ] Live-tested watched-file reload (or noted in concern if not present)
- [ ] Verdict posted via `wg log`

Depends on

• done impl-agency-scope-rules
• done .assign-peer-review-agency-scope

Required by

• done .flip-peer-review-agency-scope
• done impl-agency-integration

Log

• 2026-05-04T17:43:57.623728420+00:00 Task paused
• 2026-05-04T18:36:41.360630331+00:00 Task published
• 2026-05-04T18:36:58.527361896+00:00 Lightweight assignment: agent=Careful Programmer (f5143935), exec_mode=light, context_scope=task, reason=Careful Programmer is the best fit for a correctness-focused peer review that needs careful reading of the composer and parser behavior without making changes.
• 2026-05-04T20:23:46.093772752+00:00 Spawned by coordinator --executor claude --model opus
• 2026-05-04T20:23:56.067949858+00:00 Starting peer review of impl-agency-scope-rules — checking scope field, composition-rules.csv parser, watched-file reload, back-compat, and cap enforcement
• 2026-05-04T20:31:11.499305325+00:00 CONCUR. Reviewed src/agency/composition_rules.rs (parser+watcher), src/agency/run_mode.rs (scope helpers + filter_components_by_required_scope), src/agency/prompt.rs (scope-aware resolve), src/commands/assign.rs (apply_caps), src/commands/evaluate.rs (evaluator composer), src/agency/starters.rs (scope-tagged primitives), and tests/integration_agency_scope_rules.rs. cargo test --test integration_agency_scope_rules: 10/10 pass.
• 2026-05-04T20:31:13.955437945+00:00 Scope value set exact match: required_scope_for_task() emits task | meta:assigner | meta:evaluator | meta:evolver | meta:agent_creator (run_mode.rs:54-66); starters.rs tags primitives with the same five values. No synonyms or typos.
• 2026-05-04T20:31:26.818580382+00:00 CSV columns exact match: agent_type,rule,max_role_components,max_desired_outcomes,max_trade_off_configs,all_projects,project_ids. Parser is header-order-tolerant via case-insensitive header lookup. Empty cells -> None (unlimited). project_ids splits on ; or , and trims.
• 2026-05-04T20:31:50.327533066+00:00 Live-smoke watched-file reload via real wg binary (after cargo install --path .): ~/.agency/composition-rules.csv with 'assigner,strict,2,1,1,true,' caused wg assign --auto to log '[assign] composition-rules cap dropped 1 agent(s): 79ba3db9 (role Assigner has 7 components > cap 2)' and select Default Creator instead. Edited file to cap=10, no restart, next wg assign --auto picked Default Assigner with no drop message. Removed file, third wg assign --auto picked Default Assigner unfiltered. Confirms apply_caps re-reads via load_overlay() on every call -- no daemon restart needed.
• 2026-05-04T20:31:53.415206409+00:00 Back-compat verified: scope_matches() in run_mode.rs:74-81 treats None|Some(empty) as match-everything; test_legacy_primitives_default_to_match_all confirms untagged primitives pass any required scope. Implementation is more permissive than 'default to task' in the spec (match-all vs match-task-only), but functionally compatible -- untagged primitives still surface for both task and meta:* scopes. metadata.scope fallback honoured via component_scope/outcome_scope/tradeoff_scope helpers.
• 2026-05-04T20:32:07.195618723+00:00 Cap enforcement: integration test test_composition_rules_cap_enforces_max_role_components passes; live smoke shows over-cap composition is dropped from candidate pool. apply_caps in assign.rs:51-93 falls back to unfiltered pool (with stderr warning) if cap would empty pool entirely -- defensible 'silent failure is worse than cap violation' choice, called out in the doc comment. Note: only max_role_components is consumed for pruning; max_desired_outcomes and max_trade_off_configs are parsed but not yet enforced in apply_caps. Outcomes are 1:1 with roles so a separate lookup path would be needed -- worth a follow-up but not in scope for this task.
• 2026-05-04T20:32:25.810539267+00:00 File scope respected: git show 998a5f224 --stat lists only composition_rules.rs (new), mod.rs, prompt.rs, run_mode.rs, starters.rs, assign.rs, evaluate.rs, integration_agency_scope_rules.rs. types.rs, hash.rs, agency_import.rs untouched -- schema fields owned by impl-agency-schema-fields, hashes unaffected, CSV roundtrip untouched. Verified via git diff against base.
• 2026-05-04T20:32:32.995786675+00:00 Verdict: CONCUR. Implementation matches the validation criteria; live smoke confirms cap enforcement and watched-file reload. Two follow-up items worth tracking (not blocking): (1) apply_caps consumes only max_role_components -- max_desired_outcomes/max_trade_off_configs not enforced; (2) cap fallback-on-empty is permissive by design but a strict mode flag could be useful later.
• 2026-05-04T20:33:00.279700968+00:00 Task pending eval (agent reported done; awaiting `.evaluate-*` to score)
• 2026-05-04T20:34:21.141513847+00:00 PendingEval → Done (evaluator passed; downstream unblocks)