feat: gstack-decision-learn — calibrate /autoplan from user overrides

[HMAKT99]

Summary

• Reads Decision Audit Trail tables from past /autoplan runs
• Identifies decisions users consistently override
• Writes calibration to ~/.gstack/decision-calibration.json
• Future /autoplan runs can reference patterns to calibrate auto-decisions

$ gstack-decision-learn

Analyzed 12 autoplan runs, 184 decisions, 23 overrides.

LEARNED PATTERNS (3):
  scope_expansion_auth — overridden 4/4 (100% override rate)
  test_defer_never     — overridden 6/8 (75% override rate)
  codex_disagree_trust — overridden 5/7 (71% override rate)

Wrote: ~/.gstack/decision-calibration.json

1 file, 156 lines

bin/gstack-decision-learn — bash + inline Python. Parses existing audit trails.

Test plan

• All existing tests pass
• --show prints current calibration (or empty default)
• --reset clears calibration file
• No new dependencies

[sevastyanovio]

Cool idea — learning from user overrides is the right instinct. I'm working on something related in PR #405 (/meditate) that approaches "compound over time" from the other direction: instead of learning from explicit override decisions, it proactively scans the repo + past AI conversations (Claude, Codex, Gemini) to build a "User Taste" profile — what you correct, what you ask about repeatedly, what patterns you prefer.

These could work well together: /meditate surfaces what the user cares about, decision-learn calibrates how autoplan should respond to those preferences. Happy to coordinate if both land.

[garrytan]

Thanks for building this! Closing — decision-learn (calibrating autoplan from overrides) is an interesting idea but premature. We need more users actively using autoplan before we build a learning layer on top of it. If this becomes a real pain point we'll revisit. Appreciate the forward thinking!