phrazzld/reflect

/reflect

Structured reflection that improves both the harness and the operator.

When roster receipts exist, include .harness-kit/traces/ delegation receipts in the evidence set when they are relevant. Reflection should convert provider-lane results and failure modes into backlog, harness, or coaching outputs without inventing hidden rankings.

Delegation Floor

Delegation floor applies: probe the roster first; dispatch two or more providers for substantive work; direct solo only for mechanical, emergency, user-forbidden, or fewer-than-two-providers cases. See harnesses/shared/AGENTS.md (Roster).

Local lane guidance: Use lanes to surface independent failure interpretations and improvement proposals; the lead owns synthesis and codification choices.

Every finding becomes one of three things:

• a codified artifact
• a concrete coaching note
• an explicit justification for not codifying

Work Ledger

When .harness-kit/work/ledger.jsonl is available, /reflect consumes the latest events for the active backlog/branch plus trace refs and delegation receipts. It calls scripts/work-ledger.py append with phase_started at retro start, next_action_changed when follow-up backlog or harness proposals are emitted, and phase_completed when the reflection packet is complete. Follow-up proposals are evidence refs, not hidden chat-only state.

Routing

| Mode | Intent | Reference | |------|--------|-----------| | distill (default) | End-of-session retrospective -> codified artifacts + operator coaching | references/distill.md | | calibrate | Mid-session harness postmortem — fix the harness before the code | references/calibrate.md | | coach | Deep dive on prompt quality, technical specificity, and concept building | references/coach.md | | checkpoint | Opt-in teach-back checkpoint with restatement, verdict, gaps, and gate artifact | references/checkpoint.md | | prompt-debt | Promote repeated corrections and repeated workflow patterns into one codification proposal | references/prompt-debt.md | | tune-repo | Refresh context artifacts, detect drift, update repo guidance | references/tune-repo.md | | append | Append issue-scoped retro notes for /groom to consume later | references/retro-format.md | | cycle | Bounded end-of-ship retrospective invoked by /ship — emit backlog mutations, harness-tuning proposals, and a cycle summary for the caller to apply | references/cycle.md |

If the first argument matches a mode name, route to that reference. If no mode is provided, run distill.

Interpret natural-language requests as:

• "checkpoint this", "teach-back", "make sure I understand", or packet marker Comprehension-required: <topic> -> checkpoint
• "how could I have asked better", "teach me from this", "help me prompt better" -> coach
• "same correction again", "why do I keep repeating this", "codify this pattern", "prompt debt" -> prompt-debt
• "why did you do that", "you made the wrong call", "fix your instructions" -> calibrate
• "tune this repo", "refresh AGENTS", "context drift" -> tune-repo
• "reflect on cycle <cycle-id>", "postmortem this cycle", invocation from /ship (or transitively from /flywheel, which composes /ship) -> cycle

Responsibility Split

Reflection must separate three classes of failure:

1. Harness failure — the instructions, skills, tools, or codebase should have prevented the problem
2. Shared ambiguity — both sides left important constraints implicit
3. Operator-spec gap — the decisive information lived only in the user's head, so a tighter prompt would have reduced search space

Do not dump harness failures onto the user. If the repo, docs, or available context already contained the answer, that is not a prompt-quality critique.

Default Deliverables

Even in distill, inspect both lanes:

• System lane — instructions, skills, hooks, tests, CI, AGENTS.md, docs
• Operator lane — prompt rewrites, vocabulary, stack concepts, next-session moves

System codification is mandatory. Operator coaching is mandatory to assess, but only mandatory to emit when there is concrete, high-leverage feedback. Otherwise say so explicitly instead of manufacturing generic advice.

Use coach when the user wants the operator lane expanded into a deeper lesson.

Codification Hierarchy

When encoding knowledge, always target the highest-leverage mechanism:

Type system > Lint rule > Hook > Test > CI > Skill/reference > AGENTS.md > Memory

Prompt Debt

Prompt debt is a repeated human correction, repeated request, or repeated decision pattern that should become a durable harness artifact instead of remaining chat-only advice. Use available local surfaces only: repo-local reflect notes, review scores, delegation receipts, traces, session summaries, and durable memory notes. Chronicle-derived context may inform the pattern, but do not quote private personal detail.

When .groom/review-scores.ndjson exists and scripts/review-score-trends.py is available, run the analyzer before proposing skill changes. Treat 5+ score entries as enough for a trend; below that, report the count and avoid a tuning claim. If the analyzer names a dimension regression or high false-positive rate, propose a concrete skill/reference edit using the codification hierarchy rather than a generic observation.

Promote a pattern when it appears at least twice across sessions, or once when it prevented a shipped regression, runaway spend, data loss, or client-facing artifact error. Emit one highest-leverage proposal by default:

## Prompt Debt

- Pattern:
- Evidence count:
- Safe evidence snippets:
- Recommended target:
- Acceptance criteria:
- Residual risk:

Apply the codification hierarchy above. Prefer type, lint, hook, test, or CI coverage before skill prose; use AGENTS.md for always-on routing; use memory only for preference-level defaults that cannot be enforced.

Cycle Mode Authority (outer-loop only)

cycle is a bounded invocation: /ship calls it at the end of the final-mile pipeline to capture learnings from the just-shipped ticket. /flywheel triggers it transitively by composing /ship. When invoked as cycle, reflect gains two privileges the other modes lack:

1. Backlog mutation proposals — may propose create, edit, consolidate, reprioritize, or delete on items in backlog.d/ (never backlog.d/_done/). Every proposal must cite an evidence ref from the cycle (commit, diff hunk, receipt path, log line).
2. Harness-tuning proposals — may propose skill/agent/hook/AGENTS.md/ CLAUDE.md edits. Reflect emits; it does not apply. The caller routes these to a harness branch for human review.

All other modes are read-only against backlog.d/ and the harness. If cycle cannot cite evidence for a mutation, downgrade it to a finding and let a human decide.

Invocation Contract

Triggered as /reflect cycle (aliases: /reflect --cycle <cycle-id>). The caller — normally /ship — passes this input packet:

• branch: name of the just-shipped feature branch (pre-merge).
• merged_sha: squash commit SHA now on master/main.
• closed_backlog_ids: list of IDs closed in this cycle (the closing set from /ship's trailer scan).
• referenced_backlog_ids (optional): Refs-backlog IDs noted but not closed.

A cycle-id identifies the retro artifact; derive it from merged_sha short form when the caller does not supply one.

Output Contract

Three required categories plus one optional prompt-debt category. The structured categories must be cleanly separable so the caller can apply them under different policies.

1. Backlog mutations (structured, machine-consumable). For each:

   • action: create | edit | reprioritize | delete
   • path: concrete backlog.d/<id>-*.md target
   • body: full file content for create, unified diff for edit, new priority for reprioritize, justification for delete
   • evidence: cycle ref justifying the mutation These are proposals. The caller (/ship) applies them to master via a follow-up commit and owns commit hygiene. Reflect does not stage, commit, or push these files itself.

2. Harness-tuning proposals (structured, machine-consumable). For each:

   • path: concrete file under skills/, agents/, harnesses/, AGENTS.md, CLAUDE.md, or a hook script
   • body: unified diff or full new-file content
   • evidence: cycle ref and codification-hierarchy justification Reflect must not write these to master. The caller routes them to a harness branch (/ship uses harness/reflect-outputs). A cycle run that mutates harness files on the current branch is a bug.

3. Prompt-debt proposal (optional, structured). Include at most one by default, only when repeated corrections or high-severity prompt patterns are visible in cycle evidence:

   • pattern: short name
   • evidence_count: sanitized count, not raw transcript text
   • safe_evidence_snippets: redacted commands, file paths, or short examples
   • recommended_target: codification target from the hierarchy
   • acceptance_criteria: how the future run proves the debt is paid down
   • residual_risk: what remains ambiguous or intentionally manual

4. Cycle summary (human-readable narrative). What shipped, what was learned, what went well, what went poorly. Also written to the standard retro location (.groom/retro/<primary-id>.md or the .harness-kit/reflect/<cycle-id>/ receipts dir, matching whatever convention the invoking repo already uses).

Invariants

• Harness mutations never land on master directly. Reflect emits; the caller routes to a harness branch. This is a hard cross-skill invariant also asserted in ship/SKILL.md.
• Backlog mutations are proposals, not auto-applied. Reflect does not git add / git commit on behalf of the caller.
• Session-retrospective mode still works standalone. distill, calibrate, coach, tune-repo, and append remain usable without cycle context — cycle is additive, not a replacement.

See references/cycle.md for judgment rules (consolidate vs split, when to escalate to a harness branch, evidence standards).

Gotchas

• Blaming the user for missing repo context: If the agent could have found it, it is a harness or retrieval failure.
• Giving generic prompt advice: "Be more specific" is not feedback. Name the missing constraint, example, acceptance test, or boundary.
• Skipping the operator lane: A retro that only mutates harness artifacts leaves user leverage on the table.
• Turning coaching into scolding: Start with what the prompt achieved, then show the stronger version and why it improves the search space.
• Teaching concepts without anchoring them to the session: Vocabulary only sticks when tied to a concrete decision, bug, or design tradeoff.