Stage 7: Fix Implementation

Implement every non-discarded fix from the approved action plan.

Inputs

• eval/action_plan.md

  -- the approved action plan from Stage 5 (possibly modified during HITL in Stage 6)

• eval/stage6_decision.md

  -- if it exists, the HITL decision record from Stage 6 (contains user modifications)

• eval/baseline_metrics.json

  -- the pre-fix baseline metrics from Stage 3 (for reference in changes log)

Read the action plan and stage6 decision (if present) before starting.

Pre-flight: Git Safety Checkpoint

Before making ANY changes to source files:

1. Run

   git status

   to confirm the working tree state
2. Create a safety commit or stash:

   git stash push -m "pre-pipeline-fixes-$(date +%Y%m%d-%H%M%S)"
   

   If there are no uncommitted changes to stash, create a lightweight tag instead:

   git tag pre-pipeline-fixes-$(date +%Y%m%d-%H%M%S)
   

3. Record the stash ref or tag name in

   eval/changes_log.md

   under a "Rollback" section so the user can restore if needed

This ensures every fix is reversible with a single

git stash pop

or

git checkout

.

Pre-flight: HITL Modification Check

If

eval/stage6_decision.md

exists:

1. Read it and identify any items the user modified, added, or re-prioritized during Stage 6
2. Build a set of

   HITL_MODIFIED_IDS

   -- the insight/skill IDs that the user changed
3. When logging each fix later, tag modified items with

   [HITL-MODIFIED]

   in the changes log so reviewers know which fixes reflect user judgment vs. the original pipeline output

If the file does not exist, assume no HITL modifications were made.

Pre-flight: Conflict Scan

Before implementing any fixes, scan the action plan for potential conflicts:

1. Build a map of

   file_path -> [fix IDs that touch it]

2. If two or more fixes modify the same file, flag them as co-located
3. If two or more fixes modify the same section (within ~20 lines of each other), flag them as overlapping
4. For overlapping fixes: plan to apply them sequentially in priority order, re-reading the file between each edit to ensure the second fix still makes sense on top of the first
5. Log any detected conflicts at the top of

   eval/changes_log.md

   under a "Conflict Notes" section

Process

Work through the action plan in priority order. For each non-discarded fix:

1. Understand the fix

• Read the recommendation carefully
• Read the referenced files in the codebase
• Understand the surrounding code before making changes
• Check if this fix was flagged as co-located or overlapping in the conflict scan. If overlapping with a previously-applied fix, re-read the target file to see the current state after prior edits

2. Implement the change

For code fixes:

• Find the relevant files
• Make the minimal, targeted change described in the recommendation
• Do not refactor surrounding code unless the fix obviously breaks without light adjacent cleanup (e.g., an import is missing, a variable was renamed). If you make adjacent cleanup, log it explicitly as "adjacent cleanup" in the change entry
• Do not add features beyond what was recommended

For prompt fixes:

• Find the system prompt file (use domain context from Stage 2 if needed)
• Add the recommended instruction at the appropriate location
• Do not rewrite existing prompt text unless the recommendation explicitly says to

3. Log the change

Append to

eval/changes_log.md

:

## Fix N: [skill/insight name] [HITL-MODIFIED if applicable]
**Type:** code fix | prompt fix
**Verdict from action plan:** [quote the recommendation]
**Files modified:**
- `path/to/file.py` -- [what changed and why]
**Before:**
\```
[relevant snippet before change]
\```
**After:**
\```
[relevant snippet after change]
\```
**Linked metrics:** [which metrics this should improve]
**Conflict notes:** [if this fix overlapped with another, note it here; otherwise "none"]

4. Handle uncertainty (NEEDS REVIEW workflow)

If a fix requires changes you are unsure about:

1. Do NOT implement it
2. Log it as

   NEEDS REVIEW

   in the changes log with:
   • What specifically is unclear
   • What information would resolve the ambiguity
   • The files and lines you examined
3. Continue to the next fix -- do not block the pipeline
4. At the end of all fixes, collect all NEEDS REVIEW items into a dedicated section (see Output format below). The pipeline does NOT stop; these items are presented to the user after all other fixes are applied.

Post-Fix: Next Steps (Do NOT Re-run Baselines)

Do NOT re-run

compute_baselines.py

as part of this stage. The baseline metrics were computed against the original traces, which reflect old agent behavior. Re-running against the same traces will show zero movement for prompt-only fixes and is misleading.

Instead, after all fixes are applied, include a Next Steps section in the changes log that tells the user:

1. Generate new traces by running the agent with the updated prompts/code
2. Then re-run baselines against the new traces:

   python eval/compute_baselines.py --traces-dir <new_traces_folder> --output eval/post_fix_metrics.json
   

3. Compare

   eval/post_fix_metrics.json

   against

   eval/baseline_metrics.json

   to measure actual improvement

Rules

• Do NOT modify trace files
• Do NOT make changes beyond what the action plan recommends (except adjacent cleanup logged explicitly)
• Make minimal, targeted changes -- don't clean up or refactor surrounding code
• If the action plan says "discard", skip that entry entirely
• You MAY write

  eval/changes_log.md

  as the primary output
• Do NOT run

  eval/compute_baselines.py

  -- baselines should only be re-computed after new traces are generated with the updated agent

Output format

Write

eval/changes_log.md

:

# Changes Log

## Rollback
- **Safety ref:** `git stash` ref or tag name
- **To undo all fixes:** `git stash pop` or `git checkout <tag>`

## Conflict Notes
- [any file/region conflicts detected, or "No conflicts detected"]

## Summary
- Code fixes applied: N
- Prompt fixes applied: M
- Skipped / needs review: K
- HITL-modified items: J

---

## Fix 1: [skill name]
...

## Fix 2: [skill name]
...

---

## Needs Review
[Collected list of all NEEDS REVIEW items with context, or "None -- all fixes applied successfully"]

For each NEEDS REVIEW item:
- **Fix N: [skill name]**
- **What is unclear:** [specific ambiguity]
- **What would resolve it:** [information needed]
- **Files examined:** [paths and lines]

---

## Next Steps

To measure actual improvement:
1. Generate new traces by running the agent with the updated prompts/code
2. Re-run baselines:
\```bash
python eval/compute_baselines.py --traces-dir <new_traces_folder> --output eval/post_fix_metrics.json
\```
3. Compare `eval/post_fix_metrics.json` against `eval/baseline_metrics.json` to measure metric deltas

Outputs

• eval/changes_log.md

  -- full log of all changes, conflicts, NEEDS REVIEW items, and next steps
• The actual code/prompt changes in the repository
• A git stash or tag for rollback