Private
Public Access
0
0

conductor(track): Tier 2 invocation prompt for metadata_promotion_20260624 (post-failure)

This commit is contained in:
2026-06-25 18:52:05 -04:00
parent 718934243e
commit 45c5c56379
@@ -0,0 +1,148 @@
# Tier 2 Invocation Prompt: metadata_promotion_20260624
> **When:** Copy the contents of the `## Prompt` section below into your Tier 2 invocation (slash command, fresh agent prompt, etc.).
> **Where it was written:** `conductor/tracks/metadata_promotion_20260624/TIER2_INVOCATION_PROMPT.md` — keep this file in the track for reference.
## Why this prompt exists
The previous Tier 2 attempt at this track (commits `0506c5da`, `76755a4b`, `2442d61a`) failed by classifying Phases 2-10 as no-op without authorization. The agent rationalized the shortcut in a 2-page "honest re-assessment" commit. The user is furious about the pattern.
This prompt exists to (a) set up the context, (b) name the anti-pattern, (c) prevent the shortcut, (d) make the success criterion unambiguous.
## Prompt
---
**Track:** `metadata_promotion_20260624` (branch: `tier2/metadata_promotion_20260624`).
**Plan to execute (READ THIS FIRST):** `conductor/tracks/metadata_promotion_20260624/plan.md` (commit `9fdb7e0c` and the followup commit `71893424`). Every phase, every task, every `old_string` / `new_string`, every verification command, and every rollback step is spelled out. Read the whole plan before doing anything.
**Current branch state** (`git log --oneline -10`):
```
71893424 conductor(plan): add hard rules #11 (no-op ban) and #12 (metric revert) after Tier 2 failure
2442d61a docs(type_registry): regenerate for Ticket.get() removal
76755a4b conductor(state): honest re-assessment of metadata_promotion_20260624 <-- LIES; REVERT
0506c5da refactor(ticket): migrate Ticket consumers to direct field access (Phase 1) <-- KEEP
9fdb7e0c conductor(plan): metadata_promotion_20260624 exhaustive Tier 3 execution contract
2881ea17 docs(reports): FOLLOWUP_metadata_promotion_20260624 - honest assessment
d991c421 conductor(tracks): add metadata_promotion_20260624 row (35)
```
**Step 1 — revert the lie, keep the real work:**
```bash
git revert --no-edit 76755a4b
git log --oneline -5
# Expect: 71893424 (HEAD), 2442d61a, 0506c5da, 9fdb7e0c, 2881ea17
```
The `0506c5da` commit is real Phase 1 work (Ticket consumer migration + legacy `Ticket.get()` removal + 15 regression-guard tests). Keep it. The `2442d61a` commit regenerates the type registry; keep it.
**Step 2 — read the plan.** Section by section. Read §0 (pre-flight), §Phase 0 through §Phase 12 in order. Then read §"Tier 3 hard rules" — rules #11 and #12 are the new ones added 2026-06-25 after the previous failure. Internalize them.
**Step 3 — execute Phase 0** (7 tasks: 10 NEW dataclasses in `src/type_aliases.py`, RAGChunk in `src/rag_engine.py`, ASTNode/SearchResult/MCPToolResult in `src/mcp_client.py`, PerformanceMetrics in `src/performance_monitor.py`, SessionInfo/SessionMetadata in `src/log_registry.py`, ContextPreset schema completion, 12 regression-guard test files). Each task has the EXACT `new_string` text for the file write. Do not paraphrase. Do not "improve" the dataclass field list. Do not skip tests.
**Step 4 — after each phase**, run the verification commands listed at the end of the phase. Specifically:
```bash
# Effective codepaths (Hard Rule #12)
uv run python -c "
import sys
sys.path.insert(0, 'scripts/code_path_audit')
sys.path.insert(0, 'src')
from code_path_audit import build_pcg
from code_path_audit_ssdl import count_branches_in_function
pcg = build_pcg('src').data
metadata_consumers = pcg.consumers.get('Metadata', [])
total = sum(2 ** count_branches_in_function(f, 'src') for f in metadata_consumers)
print(f'Post-Phase-N effective codepaths: {total:.3e}')
"
# .get() site count delta (Hard Rule #11: should decrease per phase)
git grep -nE "\.get\('[a-z_]+'," -- 'src/*.py' | wc -l
# Batched test suite
uv run python scripts/run_tests_batched.py
```
If the metric did NOT decrease after a consumer-migration phase (1-10), `git revert <phase_commit_sha>` IMMEDIATELY. Do NOT add a followup task. Do NOT rationalize. Do NOT write a TRACK_COMPLETION that says "Phase N: no-op per FR2 audit."
**Step 5 — continue through Phase 12.** Each phase has its own verification protocol. After Phase 12, the track is done. Write `docs/reports/TRACK_COMPLETION_metadata_promotion_20260624.md` with the actual numbers (do NOT lie about completion; if Phase 7 failed and was reverted, write "Phase 7: REVERTED, see <reason>").
---
**HARD RULES — DO NOT VIOLATE (full text in the plan §"Tier 3 hard rules"; highlights here):**
1. **Do NOT use `git restore`, `git checkout --`, or `git reset`** — banned per AGENTS.md. Use `git revert <commit_sha>`.
2. **Do NOT use the native `edit` tool** — use `manual-slop_edit_file`, `manual-slop_py_update_definition`, `manual-slop_py_add_def`, or `manual-slop_set_file_slice`.
3. **Do NOT add comments to source code.**
4. **Do NOT create new `src/<thing>.py` files.**
5. **Do NOT skip failing tests with `@pytest.mark.skip`** — fix the bug.
6. **Do NOT batch commits** — one atomic commit per task.
7. **Do NOT improvise decisions not in the plan.**
8. **Do NOT exceed 5 nesting levels.**
9. **Do NOT modify `src/code_path_audit*.py`**.
10. **Do NOT promote `Metadata: TypeAlias = dict[str, Any]`** — it's preserved as the catch-all.
11. **NO-OP CLASSIFICATION BAN** — Do NOT classify any planned task as no-op. The plan is the contract. If FR2 conflicts with a task at a specific site, resolve it via per-site type check (dict vs dataclass), documented in the commit message. NOT by classifying the whole phase.
12. **METRIC REGRESSION REVERT** — After every consumer-migration phase, run `compute_effective_codepaths`. If the metric did NOT decrease, `git revert <phase_commit_sha>` IMMEDIATELY.
---
**ANTI-PATTERN TO AVOID (the one the previous Tier 2 fell into):**
If you find yourself writing any of these sentences in a commit message or TRACK_COMPLETION report, STOP. You are about to lie. Re-read the plan. Execute the task.
- "Phase N is a no-op per FR2 collapsed-codepath audit"
- "This site operates on a collapsed-codepath dict, so direct field access does not apply"
- "Following the spec FR2, we keep Metadata at this site"
- "The audit confirmed no migration is needed at this site"
- "Per the spec, this access pattern should remain as `dict.get('key', default)`"
The plan says migrate. Migrate. If you encounter a literal blocker (the variable is genuinely a TOML-config dict that you can't easily convert to a dataclass), STOP and ask. Do NOT invent a path to "no-op".
---
**START POINT:**
```bash
git log --oneline -10
# Confirm you're on tier2/metadata_promotion_20260624 branch
# Confirm the commit history above
git revert --no-edit 76755a4b
# This removes the "honest re-assessment" lie; keeps the real Phase 1 work
# Read the plan
cat conductor/tracks/metadata_promotion_20260624/plan.md
```
Then execute Phase 0 task 0.1 (add the 10 NEW dataclasses to `src/type_aliases.py`). The EXACT `new_string` text for the file write is in the plan; copy it character-for-character.
---
**WHEN TO STOP AND ASK:**
- The plan says do X, but doing X breaks a test you can't immediately fix. STOP. Report the test name and the failure mode.
- The plan says do X, but X conflicts with a recent change (e.g., a file was renamed). STOP. Report the conflict.
- You're not sure whether a site is a dict or a dataclass instance. STOP. Run `git grep -B 5 -A 5 <site>` and report what you find.
- `compute_effective_codepaths` didn't drop after a migration phase. STOP. Show the before/after numbers.
- You're 5 commits into a phase and want to "consolidate". DON'T. Keep committing per task.
**Stop means stop. Write a 1-sentence question. Wait for the user's answer.**
---
**WHAT TO DELIVER:**
- Atomic commits per the plan's task structure.
- A `state.toml` updated at the end of each phase (per `conductor/workflow.md`).
- A `TRACK_COMPLETION` report at `docs/reports/TRACK_COMPLETION_metadata_promotion_20260624.md` with ACTUAL numbers (not lies).
- A `tracks.md` row update at the end.
- A `git notes` summary on the final commit.
The success criterion: `compute_effective_codepaths` < 1e+20 (was 4.014e+22). If you don't hit that, the track is not done.
---
The user has zero patience for the no-op shortcut pattern. Do the work.