Files
ironclaw/docs
Illia Polosukhin 417ee611df docs(plan): update engine v2 architecture to match verified reality (#2801)
* docs(plan): update engine v2 architecture plan to reflect verified reality

The plan doc claimed several items as missing/pending that are already
implemented. Update to match ground truth so future readers don't redo
the verification pass.

Changes:
- Compaction (§4.3): marked DONE, pointer to orchestrator/default.py:240-310
- Tool reliability (§4.9): tracker exists; integration tracked in #2800 PR-B
- Routines/Jobs (§6.7): routine_to_mission_alias already translates routine_* calls;
  create_job aliasing tracked in #2800 PR-C
- Two-phase commit (§6.7): marked IMPLEMENTED via unified gate
  (policy.rs:126-169 + structured.rs:139-171); simulate/preview
  intentionally not added at policy layer
- Acceptance testing (§6.7): pointer to with_engine_v2 harness; coverage
  expansion tracked in #2800 PR-D
- Phase 7: split into 7a (engine-side, DONE) and 7b (host cleanup,
  blocked on default flip)
- Status header + Implementation Progress table: updated to match current
  state; default-flip work consolidated under issue #2800

No code changes.

Refs: #2800

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* docs(plan): address review feedback on engine v2 architecture plan

Apply accuracy fixes from PR #2801 review:

- Compaction threshold: describe as configurable via `compaction_threshold`
  (defaults to 85%), matching `compact_if_needed` in the Python
  orchestrator rather than claiming a fixed 85%.
- Token estimation: move ownership to the Python orchestrator (which
  runs the chars/token heuristic); Rust no longer claims to own this.
- Compaction cross-reference: drop the stale "crate-structure block
  above includes executor/compaction.rs" note — compaction lives
  entirely in Python.
- Reliability injection details (`ENGINE_V2_RELIABILITY_HINTS` kill
  switch, `EffectBridgeAdapter` write-backs, `build_step_context`
  reads) are labelled as proposed PR-B follow-up work rather than
  described as verified reality.
- Denylist phrasing: make it clear that `build_software` remains the
  only hard-denylisted v1 tool *after* PR-C lands, not before.
- Provenance rules: document accurately that `ToolOutput` provenance
  only injects `RequireApproval` on `Financial` effects; `WriteExternal`
  taint comes only from `LlmGenerated`, per policy.rs:126-169.
- Engine-side cleanup: acknowledge that `Session` / `Routine`
  identifiers still appear in engine docs/comments; the invariant is
  no runtime dependency, not zero string occurrences.

No code changes.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-22 19:00:47 +09:00
..