Add --budget-conscious CLI flag for session-aware experiment execution

[lukeinglis]

Closes #349

Changes

• Added --budget-conscious boolean flag to both factory ceo and factory run argparse definitions, following the same pattern as --no-github
• When set, _build_ceo_task() appends a ## Budget-Conscious Mode markdown section to the CEO task string with three throttles:
  1. Skip Reviewer agent (step 2e) for diffs under 50 lines when CEO's own review verdict is CLEAN/PROCEED
  2. Cap REDIRECT iterations at 1 per agent (instead of default 2)
  3. Defer operational execution: produce code-only PRs with execution instructions in PR description
• Added a ## Budget-Conscious Mode section to the CEO prompt (factory/agents/prompts/ceo.md) defining the protocol
• Added 8 unit tests validating parser flags, task injection, and end-to-end flag propagation through cmd_ceo and cmd_run

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 87.54%. Comparing base (f72f5e6) to head (466dff5).
⚠️ Report is 126 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #346      +/-   ##
==========================================
+ Coverage   86.97%   87.54%   +0.57%     
==========================================
  Files          51       60       +9     
  Lines        7276     9148    +1872     
==========================================
+ Hits         6328     8009    +1681     
- Misses        948     1139     +191     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[lukeinglis]

✅ Factory Review: KEEP

Verdict: KEEP
Reason: Prompt-only change adds sound Builder failure recovery protocol; no code changes, guard check clean, no eval regressions

Experiment: #1
Hypothesis: Better Builder failure recovery in CEO playbook: require git log + gh pr list checks before re-spawning a Builder after failure notification (issue #294 ask 3)

Score Comparison

Metric	Value
Before	N/A (prompt-only, no eval scores)
After	N/A
Delta	N/A
Threshold	0.8000

Guard Checks

Check	Result
eval_immutable	✅ PASS
scope	✅ PASS
git_clean	✅ PASS

Code Review Notes

• Well-structured recovery protocol with concrete git log + gh pr list checks
• OR-based decision logic correctly handles partial completion scenarios
• Placed correctly in error recovery section hierarchy
• UNVERIFIED: Behavioral effectiveness depends on CEO agent following instructions at runtime

---

Posted by Factory Reviewer

[lukeinglis]

✅ Factory Review: KEEP

Verdict: KEEP
Reason: Clean implementation: CLI flag plumbed correctly through parser/cmd_ceo/cmd_run/_build_ceo_task, CEO prompt updated with well-scoped throttle rules, 8 tests covering all paths, all 135 CLI tests pass, guard check clean, no scope violations

Experiment: #2
Hypothesis: --budget-conscious CLI flag: skip Reviewer for small clean diffs, cap redirects at 1, defer operational execution to user

Guard Checks

Check	Result
eval_immutable	✅ PASS
scope	✅ PASS
git_clean	✅ PASS

Code Quality Assessment

• Critical issues: 0
• Important issues: 0
• Minor issues: 0

Code Review Notes

• Flag flows parser → cmd_ceo/cmd_run → _build_ceo_task → task string injection correctly
• CEO prompt adds Budget-Conscious Mode section with 3 clear throttles, placed before Sacred Rules
• Builder Failure Recovery Protocol is a separate additive improvement referencing issue Subscription-plan users (Max/Pro) hit session limits mid-experiment — needs visibility + budget controls #294
• 8 new tests cover parser defaults, flag acceptance, task injection, and e2e propagation for both ceo and run subcommands
• No existing tests modified or broken (135/135 pass)
• Style follows existing --no-github pattern exactly

Test Results

135 passed in 61.54s

---

Posted by Factory Reviewer

[lukeinglis]

Closing: decomposed into separate PRs

This PR bundled 4 independent experiments into a single changeset. Each piece now has its own home:

Change	New PR	Status
Builder failure recovery protocol (exp #1)	Already on main	Merged
--budget-conscious flag (exp #2)	#399 (renamed to --lite with v2 design)	Open
Handoff command (exp #3)	#373	Open
Structlog instrumentation (exp #4)	#374	Open

PR #399 redesigns the budget-conscious concept into a proper lite mode with 40-50% token reduction (5 agents instead of 12, review pipeline reduction, Archivist consolidation, archive-only Researcher, baseline eval skip, invocation budget). Full spec at .factory/strategy/idea.md.