fix(planner): harness fixes F1–F6#32
Open
alfieprojectsdev wants to merge 6 commits into
Open
Conversation
Step 1 self-generated a state dir via tempfile.mkdtemp() in /tmp and ignored an explicit --state-dir. A session-limit restart wiped /tmp and destroyed an in-progress plan one step before the final write. - Add default_state_root() / create_state_dir() in shared/resources.py: step 1 now honors an explicit --state-dir (resumable) and otherwise mints a fresh persistent dir under <project>/.claude/planner-state/ (git-ignored), never /tmp. - Make step 1 resume-safe: it no longer clobbers an existing plan.json, so re-running it against a populated state dir preserves prior work. - Add tests/test_state_dir.py guarding the persistent-dir + resume contract end-to-end. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
QR fix mode said "fix these findings", so the fixer patched only the QR-named instances. The next verify pass then re-surfaced sibling instances of the same class (temporal comments, fragmented requirements, mis-scoped comments) -- 3+ fix<->verify iterations per class. - Add shared/fix_mode.py: CLASS_SWEEP_DIRECTIVE reframes findings as examples of a class to sweep entirely, plus temporal_scan_gate(). Wired into all five *_qr_fix.py prompt builders (plan + exec twins). - Add a deterministic temporal scanner to shared/temporal_detection.py (scan_text / scan_plan_docs) scoped to the doc/comment surface and excluding decision reasoning (avoids false-firing on "chose"). - Add `temporal-scan` CLI command: exits non-zero while any hit remains, so the plan-phase fixer cannot report PASS with siblings present. - Fix a latent bug: exec_implement_qr_fix / exec_docs_qr_fix had get_step_guidance(step) but used module_path/kwargs -> they crashed on every real dispatch. Signature corrected to match the plan-phase twins. - Add tests/test_temporal_sweep.py. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Plan-phase QR applied implementation-grade checks: it FAILed a correct
plan because the working tree wasn't migrated yet ("context still shows
pre-change state") and because @@ offsets weren't git-apply-clean --
neither is appropriate for a planning artifact whose hunks nobody applies.
- Add PLAN_INTENT_FIDELITY_NOTE in qr_verify_base.py and lead every
plan-code verify check with it.
- Reframe the CONTEXT LINE and DIFF FORMAT checks in plan_code_qr_verify:
context lines must match the CURRENT (un-migrated) file (the diff anchors
to real content); @@ line numbers and git-apply-cleanliness are advisory
and reserved for the exec/impl phase. FAIL only when the diff
mis-describes the change (wrong file/transform/context).
- Reframe the plan_code_qr_decompose STEP_1/STEP_2 prompts the same way,
listing offset/apply-cleanliness as explicit non-failures.
- Extend tests/test_qr_verify_routing.py with intent-fidelity cases.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Code changes were stored as hand-authored unified diffs, so the LLM had to compute `@@ -A,B +C,D` line arithmetic by hand -- a recurring off-by-1/2 hunk source (a sub-case of the F2 whack-a-mole it can never fully win). - Add shared/diffgen.py: build_unified_diff(file, old_snippet, new_snippet, file_text) computes the hunk headers via difflib. With the current file available the diff anchors to real line numbers; otherwise it is snippet-relative (offsets advisory per F3). The LLM never writes a @@ line. - Add old_snippet/new_snippet to the CodeChange schema (source of truth). - set-change CLI gains --old-snippet/--new-snippet: the tool reads the file and stores a computed diff; mutually exclusive with --diff. - plan_code_execute step 3 now leads with the snippet workflow and demotes hand-authored --diff to a discouraged fallback. - Add tests/test_diffgen.py (unit + set-change CLI integration). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fix<->verify loops had no per-item iteration cap; only a human deciding "override / accept advisory" broke them, so a single stubborn item could loop indefinitely. - Track per-item fail_count (cli/qr.py update-item increments it on FAIL). - Add classify_item_convergence / convergence_partition in shared/qr/utils: after QR_CONVERGENCE_CAP (default 2) FAILs on the same item, a MUST item -> "escalate", a SHOULD/COULD item -> "auto_accept". - has_qr_failures and routing.detect_qr_state are now convergence-aware: auto-accepted and already-accepted items no longer route back into fix mode; MUST items past the cap still block (and escalate). - record_auto_accepts marks SHOULD/COULD items accepted-with-rationale; the qr_route_step records and logs them before routing. - build_gate_output gains an escalation path: when a MUST item is past the cap it renders ONE human-decision gate (keep fixing vs override-accept) instead of silently re-looping. - Add `accept-item` CLI command to record an override with a rationale. - Add tests/test_convergence.py. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The architect emitted planning_context in a richer shape than the schema
(constraints as dicts not strings; rejected_alternatives missing the
required rejection_reason/decision_ref) -> raw pydantic errors that blocked
a downstream --step N until hand-coerced.
- Add validate_planning_context() in shared/schema.py: runs on the RAW dict
(works even when a full Plan.model_validate would crash) and returns
targeted, self-correctable messages.
- Add `validate-planning-context` CLI command surfacing those errors and
exiting non-zero.
- Wire it into the architect execute step 6 and plan_design_qr_fix step 3
so shape drift is caught and self-corrected in the architect step, never
blocking a later step with a traceback. Also fixes a latent literal
"{state_dir}" (missing f-string) in the qr-fix validate command.
- Add tests/test_planning_context_shape.py.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.