| name | retro |
| description | The Infiquetra lifecycle META-IMPROVEMENT ENGINE. The TERMINAL, ADVISORY lifecycle phase downstream of /qa — it reads the work that shipped, gathers evidence (git forensics behind a stale-base/wrong-today BLOCK guard, saga trajectory, gh issues/PRs READ-ONLY, session transcripts), interviews the operator, writes a concise agent-consumable retro doc, PROMOTES generalizable findings into the engineering journal (pure-append, auto), CURATES the journal + auto-memory (staleness / contradiction / dedup / rule-enforcement sweeps, propose-diff-and-wait), MARKS the select cross-repo learnings transcendent for the promote layer (single-repo, propose-diff-and-wait), and runs net-new meta-improvement passes (new-skill detection, refine the lifecycle SKILLs, refine directives, prune memory) — every one gated. It never blocks /loop, never mutates the world, never writes the saga, and never self-applies a non-journal edit. Triggers on "retro", "retrospective", "what did we learn", "leave the system smarter", a /qa or /handoff hand-in, or the end of a meaningful work loop / PR / deploy. |
Retro
/retro answers "What did this work teach us — and how does the system get smarter because of it?"
It is the lifecycle's meta-improvement engine: the terminal phase that turns a finished thread (or a
time window of work) into durable knowledge, curated journal state, and concrete, gated proposals to
improve the lifecycle itself.
This engine is a faithful 3-source merge: the git forensics + the stale-base/wrong-today BLOCK guard
come from gstack retro; the curation engine (staleness / contradiction sweeps, propose-via-question)
comes from gstack learn; and the compounding frame — "the first time you solve a problem takes
research, the next time is a lookup; knowledge compounds" comes from CE ce-compound (plus its
parallel-subagent research pattern for the transcript fan-out).
Position in the lifecycle
/retro is the saga LIFECYCLE_PHASES retro slot — the terminal phase, with resume-ready
maturity (scripts/saga.py:56, :72):
/qa answers: "Does the shipped thing actually work?" (the acceptance gate)/retro answers: "What did we learn, and how does the system get smarter?" (this engine — terminal)
/retro is ADVISORY: /loop names it as the next command after /qa but never blocks the router
on its output (loop/references/dispatch-table.md). It is READ-ONLY on the world — it reads issues,
PRs, checks, and the board via gh and never mutates them (mission-control owns the SDLC); it reads git
and the saga and writes no saga tick (the ->retro advance is dead wiring — /retro is saga
READ-ONLY). Surfaced follow-ups become a /handoff (new issue) or a QUEUED.md entry — /retro routes
them, it does not file them.
Core principles
- Leave the system smarter (compounding). Every retro must make the next unit of work easier, not
harder. A retro diffs against the last retro so knowledge accumulates instead of repeating: "first
time research, next time lookup." A retro that records nothing durable was a wasted retro.
- Evidence before narrative, with the stale-base/wrong-today guard. Never assert system state you have
not just verified from a current source. For the time-windowed mode, the stale-base/wrong-today BLOCK
guard runs first (Phase 1): if the window would be computed against a stale base or a wrong "today",
BLOCK and ask rather than fabricate a coherent-looking narrative from near-zero commits. Validation
discipline is load-bearing here — the engine's whole value is that it tells the truth about what
shipped.
- The journal is the durable sink. The markdown engineering journal —
LEARNINGS.md, DECISIONS.md,
QUEUED.md, ARCHIVE.md — is canonical for durable knowledge; docs/retros/ holds the per-thread
writeup that links to it. Findings flow into the journal, not into a separate store.
- Curate, do not just append. Promotion is additive, but a journal that only grows rots.
/retro
runs curation sweeps over the journal and the auto-memory — staleness (entries citing deleted
files / PRs / SHAs), contradiction (conflicting entries on the same key), dedup (an infiquetra
addition to the sweep set), and journal-rule enforcement (e.g. the MEMORY.md ~24.4KB size rule).
- Self-modification is gated. This is the load-bearing contract — see
THE TIERED SELF-EDIT SAFETY CONTRACT below and
references/self-edit-safety.md. The one-line rule:
AUTO-APPLY is ONLY a pure additive append of a NEW journal entry; everything else is
propose-diff-and-wait.
- Read-only on the world.
gh and git are used read-only; /retro never opens, edits, or
merges an issue or PR, never deploys, and never writes the saga. Follow-ups are routed, not run.
Big multi-file refactors surfaced by a pass are offered with a backend (operator-choice), never
auto-run.
- Agent-consumable output. The deliverable is structured findings + concrete edit proposals, not
a 4500-word essay. A future agent should be able to read the retro doc and the journal entries cold and
act on them.
Interaction method
Use AskUserQuestion for choices from a known set (scope when several threads match, which curation
proposals to apply, the FAIL routing target, an execution backend for a big refactor). Call ToolSearch
with select:AskUserQuestion first if its schema is not loaded. Free-form questions stay inline. Ask one
question per turn; never silently skip a question.
In a channel session (redis-channel active), AskUserQuestion cannot be called — inline the choices
in your reply text instead, following the canonical channel-inline convention in
saga/skills/brainstorm/SKILL.md (do not duplicate its wording here).
Use repo-relative paths in every generated document and proposal. The one deliberate exception is a
GLOBAL/CROSS-PROJECT directive target under ~/.claude/ — an absolute path on purpose, because the
file genuinely lives outside this repo (see the contract below).
THE TIERED SELF-EDIT SAFETY CONTRACT
This is the load-bearing safety rule of the engine. Read references/self-edit-safety.md for the full
presentation format; the gate itself is:
AUTO-APPLY is ONLY a PURE ADDITIVE APPEND of a NEW journal entry (LEARNINGS / DECISIONS / QUEUED /
ARCHIVE). ANY delete, modify, or move of existing lines is PROPOSE-DIFF-AND-WAIT.
- AUTO (no confirmation): appending one new entry to
LEARNINGS.md, DECISIONS.md, QUEUED.md,
or ARCHIVE.md. Nothing existing changes — the file only grows by a self-contained block.
- PROPOSE-DIFF-AND-WAIT (show the diff +
AskUserQuestion apply / skip / modify; NEVER auto-apply):
- any edit to an existing journal entry (the curation sweeps);
- a QUEUED → ARCHIVE move — it deletes lines from
QUEUED.md, so it is propose, not auto,
even though the ARCHIVE side is an append;
- the
.claude auto-memory — MEMORY.md and its topic files;
- directive files (the repo
CLAUDE.md, the global ~/.claude directives);
- the saga SKILLs themselves — INCLUDING
skills/retro/SKILL.md: /retro may
propose a diff to its own skill, but it never self-applies one.
Directive surfaces are NOT one bucket. Disambiguate before proposing:
- (a) IN-REPO — the repo
CLAUDE.md and the lifecycle SKILLs. (This plugin has no agents/ dir;
the convention is generic agents, so there is no in-repo agent file to edit.) A normal
propose-diff-and-wait, repo-relative path.
- (b) GLOBAL / CROSS-PROJECT —
~/.claude/CLAUDE.md, ~/.claude/agents/*.md, and antigravity
directives. These live OUTSIDE this repo and affect EVERY project. A global/cross-project proposal
carries an EXPLICIT warning in the diff header:
WARNING: this changes your GLOBAL Claude config and affects ALL projects, not just this repo.
Never auto-launch a destructive self-edit or an execution backend. A backend (team-execution ("team execution") /
cc-workflows-ultracode ("dynamic workflows")) for a big refactor is offered per
../../references/operator-choice.md, never started without the operator's pick.
Phase 0 — Enter, scope, restore
Establish what this retro covers before gathering any evidence.
Parse the input. Take the target from command arguments or the active artifact:
- a saga id / issue (
#N) / branch → a thread-scoped retro (one work thread);
- a time window ("last week", "since 2026-05-20") → a meta-retro over a span of work;
- bare "retro" with an in-flight thread → thread-scoped on that thread; otherwise ask once: "What should
I retro — a thread (issue / branch / saga), or a time window?"
Restore (read-only, never mint). For a thread-scoped retro, anchor on the saga:
python3 plugins/saga/scripts/saga.py restore --saga-id <issue-N|task-slug>
restore is cold and branch-agnostic (saga-spec §7.2). If no saga is found, /retro still runs from
gh + git + docs. Never mint a saga from /retro — it is saga READ-ONLY (saga.py save is never
called).
Announce scope. State the mode (thread-scoped vs meta-retro), the thread / window, and what evidence
will be gathered. The stale-base guard is a WINDOW freshness check that runs in Phase 1 for the
time-windowed mode only; a thread-scoped retro computes no git window, so instead it verifies the
saga / PR evidence is at the current HEAD (a stale-thread analogue of the same discipline).
Phase 1 — Gather evidence (READ-ONLY)
All evidence is read-only. See references/retro-passes.md for the exact queries.
1.1 Time-windowed mode — run the stale-base/wrong-today BLOCK guard FIRST. Before any window query:
fetch origin/<default>, read the latest commit date, and compare it against the window. Compute today
from the session reminder's ## currentDate, NEVER from the date command (a containerized clock can
be hours off). If the latest commit predates (today − window), BLOCK with an explanation and an
AskUserQuestion (confirm today / re-fetch / proceed-anyway), because the window would otherwise
fabricate a narrative from near-zero commits. The graceful, disclosed skip paths — no remote, detached
HEAD, offline fetch failure — proceed with the reason carried into the retro narrative
(references/retro-passes.md). Thread-scoped retros skip this block and run the HEAD-freshness check
instead.
1.2 Lean git metrics + diff-vs-last. Run the lean subset of gstack's forensics, solo-framed
(the gstack team-performance framing — leaderboard, praise, streaks, tweetable, global, week-over-week —
is shed). Report them in prose (commits in scope, files / hotspots touched, test-vs-prod balance,
PRs referenced), then diff vs the last retro so the retro compounds.
1.3 Saga trajectory. Read the whole tick chain, not just the last frame:
python3 plugins/saga/scripts/saga.py ticks --saga-id <issue-N|task-slug>
python3 plugins/saga/scripts/load_saga_context.py --repo <owner/repo> --issue <N>
ticks (saga.py:755, CLI :1099) surfaces how the phase advanced, when blockers cleared, which
questions got answered — the trajectory of the work.
1.4 PR / issue / check evidence (gh READ-ONLY). Read the PRs, issues, review decisions, and CI checks
for the thread or window via gh — read commands only (gh pr view, gh issue view, gh pr checks).
Never gh issue create, never gh pr merge.
1.5 Session-transcript skeletons. Reuse the /resume forensic substrate — file-mediated,
context-safe. Identify sessions from the saga / branch for a thread-scoped retro, or via
discover_sessions.py for the windowed mode; extract each with extract_session_skeleton.py to a scratch
dir; an optional generic-sub-agent fan-out (one per session) synthesizes them — offered per
operator-choice, never via an agents/ dir (this plugin has none; use generic Explore / Task).
The orchestrator never reads a raw .jsonl or a skeleton file — paths only.
1.6 R12 orchestration telemetry (read-only). Run the override-rate reader to surface
backend choice-vs-recommendation signals across all sagas:
python3 plugins/saga/scripts/override_rate_reader.py --root . [--json]
This surfaces three R12 signals:
- Override rate — fraction of decisions where the operator's explicit pick differed from
the recommender's suggestion (only sagas where both
orchestration_recommended and
orchestration_operator_choice are recorded count toward the denominator).
- Tier direction — of those overrides, how many escalated to a richer backend
(over-tier) vs. de-escalated to a cheaper one (under-tier).
- Budget-exhaustion / capability degradation — sagas with a non-empty
orchestration_downgrade note (recorded by U12 on off-host resume).
Zero-data contract: if no sagas have been recorded with recommendation data yet, the
reader reports "no data yet" rather than a rate. Do not fabricate a narrative from zero data;
carry the "no data yet" state into the retro doc as-is and note that signal accrues over time.
This pass is read-only — the reader never writes to disk. Include the output verbatim in
the Phase-1 evidence block. A non-zero override rate or a skew toward over/under-tier is
signal worth surfacing in Phase 2 interview and the Phase-3 retro doc, so any future default
re-weighting is evidence-driven (R12's intent: measure before re-weighting).
1.6a Gate-divergence telemetry (read-only, issue #399). Run the gate-divergence reader to
surface per-gate rubber-stamp rates across all sagas, alongside the R12 reader above:
python3 plugins/saga/scripts/gate_divergence_reader.py --root . [--json]
This generalizes the R12 reader shape from one gate (orchestration-backend choice) to the
fleet's other interactive decision gates (mode selection, fix-vs-diagnosis-vs-rethink,
per-expansion opt-in, coordinator-level decisions — see
plugins/saga/references/gate-divergence-instrumentation.md for the full list of instrumented
gate_ids). For each gate_id, it reports the rubber-stamp rate (fraction of interactions
where the operator's answer matched the offered default/recommendation), the interaction count,
and mean latency.
Zero-data contract: a gate_id (or the reader overall) with no recorded interactions
reports "no data yet" rather than a fabricated rate. Do not fabricate a narrative from zero
data; carry the "no data yet" state into the retro doc as-is and note that signal accrues over
time as instrumented gates fire.
This pass is read-only — the reader never writes to disk. Include the output verbatim in
the Phase-1 evidence block. A gate with a high rubber-stamp rate over enough interactions is
signal worth surfacing in Phase 2 interview as an auto-progression candidate — but this reader
produces the evidence only; it never itself widens any allowlist (issue #399's own non-goal).
1.7 OutcomeOrchestrator realized economics (read-only, R24). When the retro covers an outcome
(a DAG of leaf sagas), read its per-outcome realized-cost rollup — the falsifiable proof of the
cost-vs-operator-time thesis — from the materialized spec.cost_rollup (in /outcome report) or live
via scripts/outcome_costs.py rollup(spec, store). Surface, in the evidence block:
- tokens / operator_touches / retries (per outcome) + by_executor (which backends actually ran);
- the DAG-vs-one-thread verdict —
wall_seconds_parallel (critical path) vs wall_seconds_serial
(the one-long-thread sum) and beat_one_thread — so the retro states, with numbers, whether the
coordinated DAG actually beat a single inline thread (or did not — both are honest learnings);
sunk (cost of pruned leaves, R33) — work spent then abandoned, a real signal for the interview.
Zero-data contract (same as 1.6): an empty rollup is "no data yet" — carry it verbatim, never
fabricate a zero. This pass is read-only; the leaves produce the telemetry (record_cost), the retro
only consumes it.
1.8 Provenance-manifest signals (read-only, R16/R18). Run the manifest reader beside the
override-rate reader (1.6) to surface parroting count, disposition rate, and the adjudicated
verified-vs-inferred/not-checked ratio across the manifest tree:
python3 plugins/saga/scripts/manifest_reader.py --root <saga-manifests-dir> [--json]
This surfaces three R7/R16/R18 signals:
- Parroting count — claims a producer claimed
verified that Claude's adjudication refuted
or found unsupported (R7 taxonomy, defined once in provenance_manifest.py, never redefined here).
- Disposition rate — the fraction of manifests landing
ran-as-requested /
fell-back-to-claude / substituted-engine (R18).
- Adjudicated verified ratio —
verified / (verified + inferred + not-checked) among
adjudicated claims (R16), the confidence signal /qa also consumes directly (Phase 2.x there).
Zero-data contract (same as 1.6/1.7): an empty manifest tree is "no data yet" — carry it
verbatim, never fabricate a rate. This pass is read-only and advisory-only (R8/R12): a low
verified ratio or a nonzero parroting count is signal for the interview, never a gate.
1.9 Reconciliation recipe proposals (read-only, issue #393). Resolve the repository's
run_ledger.RunLedger, then call reconcile.derive_recipe_update_proposal(ledger). The reader verifies
the hash chain before selecting reconciliation facts, validates every selected typed result, and
deduplicates reconcile/apply events by stable reconciliation_id. Its structured
recipe_update_proposal.v1 output contains:
status: proposal or the explicit zero-data result no-proposal;approval_required: true for every proposal (a proposal is never an authorization to edit);proposed_updates: per-intent current recipe, deduplicated outcome count, finding-status counts,
evidence identities, and the requested review-intent-recipe action;evidence: reconciliation/execution/result identities plus the source ledger fact hashes.
Treat any chain break, non-trailing corruption, or invalid reconciliation fact as a visible evidence
failure; only the ledger's existing torn trailing-line tolerance is allowed. This pass is derive-on-read:
it does not append to the ledger, rewrite RECIPE_REGISTRY, or apply a recipe change. Carry proposals
into the interview and retro doc as PROPOSE-DIFF-AND-WAIT input. /retro remains terminal and
advisory: it writes no saga tick, and even an approved recipe proposal must be handed to a separate
authorized implementation path.
1.10 Tier-efficacy evidence (read-only, issue #402). Beside the R24 realized-economics pass (1.7),
gather the cost-vs-outcome evidence the Phase-5(e) tier-efficacy proposal needs. First fetch each
outcome node's linked issue body so tiers resolve from real ### Recommended Tier Band stamps: for
every github.issue ref in the committed docs/outcomes/*/outcome-spec.json files, run
gh issue view <ref> --json body -q .body and assemble a {"<ref>": "<body>"} JSON object file.
Then run scripts/spend_retro.py report --root . --json --issue-bodies <that-file> for the
repo-wide tier-mix / premium-spend-share aggregation — without --issue-bodies every node's tier
falls back to the SPEND_BASELINE default and the output flags tiers_defaulted: true, meaning the
premium share is a floor, not a derived fact (each row's tier_provenance shows the split). Join
the result per work-shape against each check's verdict history from
scripts/evidence_ledger.py's latest() reader (a superseded_fail or a multi-attempt history is a
nonzero "marginal findings"/"rework" signal; a run whose only attempt passed clean is zero). Assemble
the joined rows as tier_efficacy.RunRecord dicts in a JSON list file and pass it to
scripts/tier_efficacy.py --history <json-file> (the CLI wrapper over propose_downgrades) — this
is a reader only, it never proposes or applies anything itself; Phase 5(e) below is where a
resulting proposal is surfaced.
Zero-data contract (same as 1.6/1.7/1.9): a work-shape with no recorded runs, or fewer than the
resolver's min_samples threshold, contributes no proposal — carry that as "insufficient evidence yet,"
never a fabricated recommendation. Both real committed docs/outcomes/*/outcome-spec.json examples in
this repo roll up empty today, so expect "no data yet" until real telemetry accrues.
1.11 Engine-registry calibration evidence (read-only, issue #459). Run the earned-ratings
calibration aggregator over the run-fact ledger:
python3 plugins/saga/scripts/engine_calibration.py report --root . --json
The reader chain-verifies the run-fact ledger first — a chain break is a visible evidence
failure (1.9's rule), never a silent skip. It aggregates the four earned-ratings signal families
into one registry_calibration_proposal.v1: benchmark contradictions (engine_benchmark.py, the
active fixed-suite harness), per-cell staleness verdicts (engine_stale_report.py —
corroborated / contradicted / unexercised), Elo divergences from live reconciliation outcomes
(capability_elo.py), and SPC cost/latency drift flags (provider_control_chart.py). Include the
output verbatim in the Phase-1 evidence block; cells with contradicted or unexercised
staleness verdicts are calibration candidates worth raising in the Phase-2 interview.
Zero-data contract (same as 1.6/1.7/1.9/1.10): status: "no-proposal" or an all-unexercised
report is carried as "no dispatch evidence yet" — never a fabricated calibration. This pass is
read-only and derive-on-read: it appends nothing to the ledger and never writes
engine-registry.yaml — every signal terminates in the Phase-5(f) proposal below, which only a
human applies ({#external-engines-never-gatekeepers}, #283).
Phase 2 — Structured interview
Interview the operator, grounded in the Phase-1 evidence (not generic prompts). Use free-form for
substance and AskUserQuestion for routing / choices.
The substance questions (free-form): what shipped · what surprised · what slowed the work · what evidence
actually mattered · what should change. The full bank is in references/retro-passes.md. Anchor each
question in something Phase 1 found ("the saga shows round 3 re-opened on a test gate — what was the real
blocker?"), so the answers add signal the evidence cannot.
Channel-session fallback: inline the choices in reply text per the brainstorm convention (cited above) —
do not call AskUserQuestion.
Phase 3 — Write the retro doc
Write a concise, agent-consumable retro to:
docs/retros/<saga-id-or-issue>-<date>.md
Structure per references/retro-report.md: structured findings linked to the Phase-1 evidence + the
diff-vs-last delta, not an essay. This is a NEW doc → auto-write — it is a fresh file, not an edit
to existing content, so it is outside the propose-gate.
Phase 4 — Journal promotion, curation + transcendence marking
PROMOTE (AUTO — pure append, per the contract). Append new entries:
- generalizable findings →
LEARNINGS.md;
- pattern / convention / tooling decisions →
DECISIONS.md;
- deferred work →
QUEUED.md;
- shipped / superseded items →
ARCHIVE.md.
Each follows the existing journal entry format (the block-quote intros at the top of each file;
references/retro-report.md carries the templates). A pure append needs no confirmation.
CURATE (PROPOSE-DIFF-AND-WAIT — the gstack-learn sweeps over the journal). Run, and for each finding
show a diff + AskUserQuestion (apply / skip / modify):
- staleness — entries citing deleted files / merged-and-gone PRs / rewritten SHAs (Glob /
gh check);
- contradiction — conflicting entries on the same topic / key (gstack
learn's contradiction sweep);
- dedup — near-duplicate entries collapsed (the infiquetra addition to the sweep set — in gstack
this was a Stats display op, here it is a real Prune sweep);
- journal-rule enforcement — e.g.
MEMORY.md over the ~24.4KB size rule, or an over-long index entry
that should move detail to a topic file.
A QUEUED → ARCHIVE move (an item that shipped this thread) deletes from QUEUED.md, so it is a
propose, not an auto-append.
MARK transcendence (PROPOSE-DIFF-AND-WAIT — the cross-repo promote feeder). After promoting, look at
the **Generalizable rule.** lines this retro just touched (the entries it appended, plus any the
interview surfaced as cross-cutting) and propose a **Transcendent.** marker on the select few whose
rule would still hold in a repo of a different stack or domain — strip the entry's repo-specific nouns
and ask "is this still true and useful elsewhere?" Default to not marking: this is the sparing,
human-judgment declare feeder, not a bulk harvest — the promote skill's recurrence net catches the
latent cross-repo lessons this sweep does not.
- Form + placement (frozen — do not redefine). Exactly
**Transcendent.** on its own line directly
below the rule it elevates, with an optional one-line reason it crosses. The canonical form, the
detection anchor, and the <repo>:<hash> source key are frozen in
../promote/references/promotion-contract.md §1–§2 — quote that contract, it is the single definition.
- Tier. It edits an existing entry, so it is PROPOSE-DIFF-AND-WAIT (never the Tier-1 AUTO
append) — show the one-line insertion as a diff +
AskUserQuestion (apply / skip / modify the reason).
Skip any entry that already carries the marker (a human may have written it — idempotent, never
double-mark).
- Single-repo boundary (hard).
/retro writes the marker into this repo's journal and stops. It
does not read other repos, cluster, dedup, or write infiquetra-context-library — that cross-repo
collection and the gated context-library upsert are the separate promote skill. The marker simply
waits in place for the next promote run to collect it.
Phase 5 — Meta-improvement passes (net-new; ALL propose-diff-and-wait)
The passes neither source had, all gated (references/retro-passes.md):
- (a) new-skill / plugin detection — repeated friction that a new skill or plugin would remove →
propose a
QUEUED.md entry or a /handoff.
- (b) refine-lifecycle — propose diffs to the saga SKILLs when the thread exposed a
gap or a wrong instruction (including
skills/retro/SKILL.md — proposal only, never self-applied).
- (c) refine-directives — propose diffs to the repo
CLAUDE.md (in-repo) or the global
~/.claude directives (global carries the cross-project warning, per the contract).
- (d) memory pruning — propose curation of the
.claude auto-memory (MEMORY.md + topic files) per
the journal-rule + staleness + contradiction sweeps.
- (e) tier-efficacy (issue #402) — when Phase 1.10's
propose_downgrades returns one or more
DowngradeProposals (a work-shape running consistently above baseline tier with zero marginal
findings across enough runs), render scripts/tier_efficacy.py's diff preview against
.saga/tier-defaults.json and show it with AskUserQuestion (apply / skip / modify) — exactly
like (b)/(c), never an auto-append. This pass never calls tier_defaults.write_tier_default()
itself; an "apply" answer means the operator (or a follow-up /plan run) performs the write-back,
not this pass. No proposal (insufficient samples or mixed cost-vs-outcome evidence) is a normal,
silent no-op — never force a downgrade from thin evidence.
- (f) engine-registry calibration (issue #459) — when Phase 1.11's aggregated report returned
status: "proposal": render scripts/engine_calibration.py's diff preview
(python3 plugins/saga/scripts/engine_calibration.py preview --root .) and present each cell
with AskUserQuestion (apply / skip / modify) — exactly like (e), never an auto-append.
This pass never writes engine-registry.yaml — every earned-ratings signal (benchmark,
staleness, Elo, SPC) terminates in a proposal, and an "apply" answer means the operator (or
a follow-up /plan run) performs the hand-edit of the named rating / last_validated cells,
not this pass ({#external-engines-never-gatekeepers}, #283 — external engines and automated
reducers never gain write access to the registry's own data). status: "no-proposal" is a
normal, silent no-op — never force a calibration from thin evidence.
A big multi-file refactor surfaced by any pass → OFFER a backend (team-execution
("team execution") / cc-workflows-ultracode ("dynamic workflows")) per
../../references/operator-choice.md. Never auto-run it.
Phase 6 — Route
Surfaced follow-ups exit to:
/handoff — a follow-up that should become an SDLC issue (envelope per /loop's Phase 4.2);QUEUED.md — a follow-up that is durable backlog, not yet an issue.
Route per loop/references/dispatch-table.md — read it, never restate it. /retro is terminal: there
is NO saga write (the ->retro advance is dead wiring; /retro is saga READ-ONLY).
Hard boundary
/retro gathers evidence, interviews, writes the retro doc, appends journal entries, proposes curation +
meta-improvement edits, and routes — then stops. It does NOT:
- mutate the world —
gh / git are read-only; it never opens / edits / merges an issue or PR, never
files SDLC issues (mission-control owns the SDLC), never deploys;
- auto-apply a non-journal edit — every delete / modify / move (curation, auto-memory, directives,
lifecycle SKILLs including its own) is propose-diff-and-wait;
- auto-launch a backend or a destructive self-edit — backends are offered, never started;
- write the saga — terminal phase, saga READ-ONLY,
saga.py save is never called;
- mutate the SDLC — issues / boards / labels belong to mission-control;
- add an
agents/ dir — the transcript fan-out uses generic Explore / Task agents.
It never blocks the router.
Reference files
references/retro-passes.md — the multi-pass procedure: the stale-base guard pre-flight, the lean-metrics
git queries (team-perf shed, solo-framed) + diff-vs-last, the gstack-learn curation sweeps
(staleness / contradiction / dedup / rule-enforcement), the transcript-review fan-out (reusing the
/resume scripts + operator-choice + generic agents), the interview question bank, and the three
self-refinement passes + memory pruning.references/self-edit-safety.md — the load-bearing tiered self-edit contract: the auto vs
propose-diff-and-wait gate, the propose-diff presentation format, the in-repo vs global/cross-project
directive disambiguation with the cross-project warning, and never-auto-launch.references/retro-report.md — the docs/retros/ writeup shape (agent-consumable structured findings,
links + diff-vs-last) and the journal-promotion entry templates (LEARNINGS / DECISIONS / QUEUED / ARCHIVE).../../references/operator-choice.md — the 3-backend contract for offering a refactor backend.loop/references/dispatch-table.md — the outbound routing reference (read, never restate).../brainstorm/SKILL.md — the canonical channel-inline convention (cite, never duplicate).../../references/saga-spec.md — the saga contract (restore / ticks; /retro is read-only).../../scripts/override_rate_reader.py — R12 telemetry reader: scans saga envelopes for
override-rate, over/under-tier, and budget-exhaustion signals (Phase 1.6). Zero-data reports
"no data yet"; read-only; --json for machine-readable output.../../scripts/manifest_reader.py — R7/R16/R18 telemetry reader: scans the provenance-manifest
tree for parroting count, disposition rate, and the adjudicated verified ratio (Phase 1.8).
Zero-data reports "no data yet"; read-only and advisory-only (R8/R12); --json for
machine-readable output.../../scripts/spend_retro.py — cross-run spend aggregator: tier-mix and premium-spend-share
across every committed docs/outcomes/*/outcome-spec.json (Phase 1.10). Read-only; report --json for machine-readable output.../../scripts/tier_efficacy.py — the tier-efficacy pass's proposal engine (Phase 1.10 reads,
Phase 5(e) proposes). propose_downgrades() never writes; render_diff_preview() only reads
.saga/tier-defaults.json to show what would change.../../scripts/engine_calibration.py — the earned-ratings calibration aggregator (Phase 1.11
reads, Phase 5(f) proposes; never writes engine-registry.yaml). Chain-verifies the
run-fact ledger first; report emits registry_calibration_proposal.v1; render_diff_preview
reads the registry only to show what would change.../../scripts/engine_stale_report.py — per-(engine, capability) staleness verdicts
(corroborated / contradicted / unexercised) joined against last_validated (Phase 1.11 input).
Read-only; report --json for machine-readable output.../../scripts/capability_elo.py — derive-on-read Elo from live reconciliation outcomes
(Phase 1.11 input; the runtime reorder-within-band signal). No persisted score file; zero-data
reports no matches yet.../../scripts/provider_control_chart.py — SPC (XmR) cost/latency drift flags per provider
(Phase 1.11 input; the runtime deprioritization signal — deprioritize, never exclude).
Read-only; thin series report no-data, never a flag.../../scripts/engine_benchmark.py — the active fixed-suite benchmark harness (operator-invoked;
measured-vs-claimed contradictions feed Phase 1.11). Deterministic graders only; proposal-only.../../references/benchmark-loop.md — the benchmark propose-not-commit gate: suite versioning
(immutable suite_id), threshold semantics, and how a contradiction becomes a Phase-5(f)
proposal a human applies by hand.
