Decision trace

The decision-trace record, and run 115 as its worked example.

Decision trace

Composition is chatty. The resolver records what it picked and why as each decision occurs, into trace.json inside the bundle. Reasons are never reconstructed after composition. A trace is an ordered list of decisions. Each decision names the thing being decided - a personality skill, an instruction, a delivery entry point - the source it came from, the outcome, and a human-readable reason. The v0.1 outcomes are:

• selected - policy admitted the candidate.
• excluded - policy considered and rejected the candidate.
• shadowed - an identical higher-precedence copy already filled the slot.
• delivered - the adapter placed selected content at a bundle entry point.

The trace also retains one provider report for every selected provider and every configured provider excluded from the active role. Reports classify the selected person package, ordinary catalogues, and role-only skill-provider repositories, then record their configuration scope, outcome, and reason. When a role provider carries an ordinary-skill selector, its report also records the configured selector and the admitted catalogue fraction. Skills outside that slice remain explicit excluded decisions, so describe --why can distinguish selector filtering from provider or role exclusion. Each provider report carries a context-budget contribution. skills counts canonical selected skill trees attributed to that provider. context_bytes is their exact retained byte count, and approximate_tokens uses ceil(context_bytes / 4). Shadowed copies do not contribute twice. Excluded providers record explicit zero values for all three fields. Native and staged projection preserve the same trace, so the budget does not depend on the consumer layout. Selector-backed provider budgets count only the admitted slice. Invalid input fails composition with diagnostics from the in-progress trace, and no bundle is produced. Reasons are plain sentences, safe to show in a terminal and safe to keep in a public bundle. A private overlay is referenced by its source id; its content never appears in a reason. Runtime noise - durations, cache hits, terminal styling - stays out of the trace. agent-compose describe renders provider outcomes, context budgets, and the stored decisions in scannable sections, describe --why <item> follows one item to its outcome, and diff compares two bundles by decision subject plus manifest logical content ID and digest. Artifact-level changes remain visible beside logical changes. These commands do not reopen authoring roots. trace.json itself is the decision machine-readable surface; there is no second explanation format. Human output and TTY styling are views over the trace and never enter model instructions; redirected output is plain and deterministic.

Issue-suite run journal

Contract: Forgejo issue #115. Program counter uses structured bullets because the repository voice forbids prose tables.

• reconcile-current-main
  • status: completed
  • attempts: 1
  • evidence: v1.32.0, 15 open issues, baseline ward exec test and ward exec smoke passed
• role-skills-and-qa-fixtures
  • status: completed
  • attempts: 1
  • evidence: canonical role skills, compact startup cards, QA fixture doctrine, Engineer and Ops boundary, ward exec test, and ward exec smoke pass; commit cd99a44 landed; #121, #116, and #128 closed
• ceo-portfolio-operating-model
  • status: completed
  • attempts: 1
  • evidence: commit ef75fa4 landed and issue #129 closed
• close-landed-foundations
  • status: completed
  • attempts: 2
  • evidence: profile and library graph, aliases and affinities, generalized seats, copy contracts, complete evaluation matrices, and executable external profile covered by ward exec test and ward exec smoke
• complete-catalog-and-export-contracts
  • status: completed
  • attempts: 2
  • evidence: rich text and JSON catalogues, logical content manifest and diff, deterministic verified export, ward exec test, and ward exec smoke
• refresh-evaluation-baselines
  • status: completed
  • attempts: 3
  • evidence: 44 frozen responses, independent QA scores, 11 v2 records, explicit retry provenance, ward exec test, and ward exec smoke
• cross-repo-fixture-enforcement
  • status: blocked
  • attempts: 1
  • evidence: Ward #1617 landed at 2d58c641; AOS fixture enforcement landed at 8e5c7a75; live proof awaits infrastructure #685 and AOS #781
• zero-open-audit
  • status: pending
  • attempts: 0
  • evidence: none

Append-only observations

• 2026-07-28 - Current main already contained partial implementations for issues #110 through #127, but their acceptance contracts remained open.
• 2026-07-28 - The first genuine blocker was #121. Long-form role doctrine still lived in KDL and global startup eagerly embedded every role.
• 2026-07-28 - Role-skill migration reduced the unassigned Codex roster by 54%, the unassigned Claude roster by 56%, and assigned Engineer startup instructions by 70% while retaining selected skill bodies.
• 2026-07-28 - Commit cd99a44 landed the role-skill checkpoint and issues #121, #116, and #128 closed with acceptance evidence.
• 2026-07-28 - Issue #129 arrived after the initial inventory. Its CEO operating-model contract became an additional role-skill slice.
• 2026-07-28 - Foundation acceptance auditing found real partial work in #124 through #127. The checkpoint completed copy-contract provenance, whole-matrix replacement, rich catalogue JSON, export verification, and logical-content diff instead of closing the issues from commit history alone.
• 2026-07-28 - QA scored the refreshed corpus 19/44 overall: frontier 19/22 and OSS 0/22. Current qwen3:4b responses exposed reasoning, truncated, and sometimes invented facts, so the records preserve those failures.
• 2026-07-28 - Cross-repo repository work is landed. The remaining fixture proof requires an Ops-owned registry bootstrap tracked by infrastructure #685, followed by the QA fixture run tracked by AOS #781.