Finish PR #18 Codex fixes for Knowledge Loop

[Svaag]

Context

Knowledge PR #18 is the Knowledge Loop agent implementation:

• PR: Add Knowledge Loop agent #18
• Current head: 82a014fe0804d32f344d73db3090dc3d6d2fb8c6
• Merge state when captured: CLEAN
• CI/validate when captured: SUCCESS
• Local validation passed before the latest Codex findings:
  • uv run ruff check src tests
  • uv run mypy --strict src
  • uv run pytest (86 passed)
  • uv run hyrule-knowledge validate okf
  • uv run hyrule-knowledge quality --check
  • uv run hyrule-knowledge export --check
  • uv run hyrule-knowledge eval --check
  • uv run hyrule-knowledge ledger --check
  • uv run hyrule-knowledge ledger lifecycle --check
  • uv run hyrule-knowledge scan-secrets okf exports reports evals ledger schema

The Codex review cycle is not done yet; no root-post 👍 has been observed.

Remaining Codex findings to fix

1. Keep idle loop runs from publishing timestamp-only refreshes

Thread: #18 (comment)

Codex notes that the unconditional ingest step can rewrite generated concepts/exports with fresh timestamp/run metadata (last_verified_at, manifest run_id, claim timestamps) even when source snapshots are unchanged. _git_dirty() then sees a dirty worktree and --create-pr can open timestamp-only refresh PRs, making the intended idle path unreachable in real loop runs.

Suggested direction:

• Ensure no-op refresh cycles do not publish timestamp-only diffs.
• Options may include detecting/reverting timestamp-only generated/export diffs, adding deterministic/no-timestamp ingest mode, or comparing semantic/source-backed output before publish.
• Add regression coverage proving an unchanged ingest cycle can report idle and does not publish.

2. Resolve learning-event paths against the repo

Thread: #18 (comment)

Codex notes that with --repo-path set and a relative --learning-event, budget preflight currently counts events relative to the supervisor cwd, while ledger import later interprets the same relative path with cwd=repo. This can count a different file, fail before import, or make budget/import behavior inconsistent.

Suggested direction:

• Resolve relative learning-event paths against config.repo_path before budget preflight and import.
• Pass the resolved paths consistently to both load_learning_events(...) and ledger import.
• Add regression coverage for invoking the loop from outside the repo with repo-relative --learning-event.

Already fixed in this review cycle

Codex findings already addressed and resolved on PR #18:

• Ignore loop state before checking for knowledge changes.
• Persist provider budgets even when later steps fail.
• Count learning events, not input files, for import budgets.
• Charge OpenRouter attempts before child validation can fail.
• Acquire the loop lock atomically.
• Create loop PR branches from the base branch.
• Gate PR budget only for publishing cycles.
• Reset back to the base branch between loop runs.
• Charge learning-event budget before imports can partially succeed.
• Do not evict live lock holders after TTL.
• Configure a commit identity for the loop.
• Guard blank loop budget environment defaults.

Done criteria

• Fix both remaining Codex findings.
• Add/update regression tests.
• Run full repository validation.
• Amend/push PR Add Knowledge Loop agent #18.
• Resolve fixed Codex threads.
• Request another @codex review.
• Poll until either new actionable threads appear or Codex gives 👍 on the PR root post.

[Svaag]

Done. All Codex findings are fixed and merged in PR #18 (merge commit 0b414ae):

• Idle cycles no longer publish timestamp-only refreshes; learning-event paths resolve against the repo (dcda960)
• PR body reflects skipped validation (98d3247)
• schema is staged on publish; volatile detection scoped to generated metadata so a corrected event_time is preserved (725a31c)
• Failed cycles count against the daily budget; blank HYRULE_KNOWLEDGE_LOOP_STATE_DIR falls back to the cache path (f1c95fd)

Codex gave a 👍 on the PR root post with zero unresolved threads, CI green. Closing.