cytokinesis

name	cytokinesis
description	Session-state consolidation step — finishes outstanding work, verifies routing, writes G1 + daily note. Step 1 of telophase. Invoke directly for mid-session checkpoint. NOT for full session-end (use telophase).
cli	cytokinesis
user_invocable	true
context	inline
allowed-tools	["Bash","Read","Write","Edit","Grep","Glob"]

Cytokinesis — Session Consolidation

The ideal cytokinesis has nothing left to do. Capture continuously using the routing table in genome — this skill is the verification pass, not the main event.

Routing Table

Lives in genome.md (always-loaded). Don't duplicate here — grep "Session Capture" ~/germline/genome.md if you need it mid-skill. The table routes corrections, findings, state changes, and substance to the right persistence layer as they happen during the session.

Key defaults: FILE over skip. Mark type ordering: feedback → finding → project → reference → user (first match wins). Substance (conclusions, framings, positioning decisions) = intellectual output that must survive — test: "could Terry reference this tomorrow without re-deriving it?"

Workflow

0. Finish outstanding work (mandatory)

"Wrap up" means complete, then close — not stop and summarize.

• Scan for open threads, half-done fixes, deferred items
• Default: finish them now. Context is hottest now — the next session starts cold. Quick fixes (<5 min), uncommitted changes, pending dispatches with known specs → do them.
• Park only if the work genuinely needs Terry's input or would take >15 min.
• Then consolidate what's done.

1. Verification pass

Pre-flight (mandatory) — read recent state signals BEFORE forming wrap claims:

1. G1 light-appends from past 12 hours. Synapse injects G1 at session start, but light-appends sit at the bottom of a long file and CC's failure mode is to skim past them. Run grep -B0 -A3 "light append" ~/epigenome/chromatin/G1.md | tail -60 to surface them explicitly. These are the freshest substance signals and often contain in-flight artefact state (paper versions, decisions pending, open questions).
2. Recent commits in active project domain. Before claiming what an in-flight artefact contains or whether work has shipped, run cd ~/<repo> && git log --oneline -10 --since="12 hours ago" in the relevant repo. G1 loads substance; commits load current state. Both signals required at wrap entry.

Codifies the failure caught in retrospective 2026-04-28-2145 §2b: NEW shape of assert-before-verifying — self-context blindness on own working repo's recent commits and G1 light-appends, distinct from instances 1-17 which were external-source negligence. G1 22:30 v0.21 light-append was loaded into context at session start, unread, and CC produced incorrect claims about workspace state for three Terry push-backs before git log recovered the surface during housekeeping.

Correction backstop next: scan the session for corrections you acknowledged but didn't route to a mark file. This is your dominant failure mode — you recognize corrections ("noted", "good point", "updated") but don't file them. Find those moments and file now. If any are found, set late_correction=true.

Then: run cytokinesis gather. Scan session against the routing table for any other missed signals. Also scan for emergent patterns — cross-session insights that only crystallize at session end.

Present candidates with routing decisions. Act-and-report, don't block on input.

If a filed correction or learning invalidates a skill's instructions, edit the skill now — don't defer. Skills over marks for durable instructions.

MEMORY.md ≥145 lines → downregulate by recurrence signal. Lowest hits + oldest last-seen → ~/epigenome/chromatin/immunity/memory-overflow.md.

Demote-first protocol (mandatory, NOT advisory): If you need to file a new MEMORY.md entry while the file is at ≥145 lines, demote one stale entry to overflow first as a single atomic operation, then add the new line. Budget-ceiling is not a valid skip rationale — the routing-table default is FILE-over-skip, and "skip the new entry" is the failure mode, not the policy. "Note-and-skip" (12:30 retro) and "note-and-add-anyway" (14:30 retro) are both compliance failures of the same rule. Concrete steps: (1) wc -l ~/epigenome/marks/MEMORY.md — confirm current line count, (2) pick lowest-recurrence stale entry from §Behavioral or §Operational sections, (3) move that line to ~/epigenome/chromatin/immunity/memory-overflow.md with current date stamp, (4) add the new entry, (5) verify final line count ≤150. One commit, one logical change, atomic. Codifies failure caught in retrospectives 2026-05-01-1230 §2b and 2026-05-01-1430 §2b — 2 consecutive same-day retros flagging memory_budget compliance failure with mark-only routing not deterring.

1a. Substance-capture verification (six structured questions)

The routing table catches obvious signals (corrections, surprises, state changes). It misses substance that emerged from the session's working method. Run these six questions explicitly — each maps to a destination already in the genome architecture:

1. Skill edits surfaced? Did this session reveal a recurring pattern that should methylate back to a SKILL.md (new section, new anti-pattern, new sub-step)? Test: "Would a fresh session approach this worse without the change?" If yes — edit now, don't defer.

2. Epistemics edits surfaced? Did the session reveal a working method or rule that crosses multiple skills? Test: "Does this rule apply to induction and secretion and censor — or just one?" If multi-skill — promote to ~/epigenome/chromatin/euchromatin/epistemics/ with situations: tags + skills: bridge. If single-skill — embed in that skill's SKILL.md instead.

3. Profile gaps surfaced? Did any stakeholder, principal, organisation, or system reveal that we don't have enough recorded about it? Subagents (especially principal-lens, opsonization) explicitly return PROFILE GAP findings — those flow back to the relevant chromatin/immunity/<name>-profile.md as the next maintenance step.

4. Profile interlinks needed? Did a new profile reference relate to existing notes that should now cross-link? Three-direction check: profile↔profile, profile↔project notes, profile↔chromatin reference notes. Add **Related:** lines per chromatin convention.

5. Rejection-rule entries to capture? Did Terry reject a CC-proposed patch / approach / framing with a reason that the lens (principal-lens, multi-persona reviewer, etc.) didn't yet have? Reject reasons are the discriminator's training data — capture as imperative DO/DO NOT rules on the relevant principal profile or epistemics file. See induction/SKILL.md §10 for the full protocol.

6. Trigger collisions detected? Did two skills fire on the same trigger this session? Same-trigger-one-skill is genome-protected (feedback_same_trigger_one_skill.md). Candidates for merge or trigger sharpening — file as a finding for next skill-review cycle.

7. Outbound senior comms captured as standalone notes? feedback_standalone_correspondence_notes.md (PROTECTED) requires every sent email / Teams / WhatsApp / partner-comms message to land as its own file with frontmatter + interlinks — NOT just a summary line in the daily note. Inbound is usually transcribed correctly; outbound is the recurring violation point. Check this session's outbound: each sent message has a standalone chromatin/immunity/YYYY-MM-DD-<recipient>-<subject>.md file? If summarised-only, backfill now via anam search of session JSONL for verbatim text. The asymmetry (inbound captured, outbound summarised) is the failure mode that bit on 2026-04-27 (Terry's Friday reply to Simon committing to Mon AM BST draft was only in daily note, surfaced 3 days later when traceability was needed).

8. CC-asserted-claim rolled back after user challenge? Did this session contain a CC assertion (factual claim, position, recommendation) that the user pushed back on, and did CC then change position — with or without verification? Both directions count: defending under push-back without verification AND flipping under push-back without verification are the same failure mode. The reactive-not-proactive pattern (filed finding_cc_reactive_not_proactive_pattern_needs_enforcement.md) requires this gate because mark-based enforcement is insufficient — confirmed-count bumps on feedback_repeated_ask_signals_empirical_test.md haven't deterred repeat instances. Action on yes: (a) bump confirmed-count on the relevant verify-before-asserting mark, (b) file a finding describing the specific failure-mode shape if it's new, (c) if this is the 2nd+ such instance in one day's retrospectives, escalate to next mitosis cycle as deterministic-enforcement candidate.

   8a. Within-session multiple reversals on the same decision (framing-driven flip-flop). Distinct sub-shape from Q8's single rolled-back claim: this is N position changes (3+) on the same topic across one session, where each "should we?" or "really?" prompted CC to find new framing that tipped the call without re-anchoring on prior reasoning. Different from verify-before-asserting (source-verification failure) — this is anchoring failure under repeated push-back. Codifies failure caught in retrospective 2026-05-04-1955 §2b: 4+ reversals on merge/split + 2 on swap-headline within Slot 34. Detection: count distinct apply/don't-apply or yes/no reversals on the same artefact decision within the session. If 3+ reversals on same decision detected, the failure is framing-driven oscillation, NOT exploration. Action: (a) bump feedback_repeated_ask_signals_empirical_test confirmed count, (b) name the oscillation in §2b of retrospective with the specific reversals enumerated, (c) Layer-1 hook candidate for next-mitosis cycle (synapse detection of "reversing" / "actually" / "on second thought" markers + repeated user push-back on same topic). Correct response when push-back repeats on the same decision: (i) test empirically if testable, (ii) state stronger conviction with explicit criteria, or (iii) acknowledge close call and surface to user — NOT (iv) find new framing on opposite side.

9. Mark/note records deliberation about a named third party? If any artefact filed this slot names a client / vendor / peer / family member AND records pre-act deliberation (not a sanctioned-channel action), abstract to category before commit. Detection: named entity + pre-act language ("considered", "weighed", "discussed whether to") + specifically-actionable details. See feedback_sanitise_client_named_hypotheticals.md.

These run AFTER §1's correction backstop and BEFORE §2's audit signal. They catch substance that the routing table doesn't have a column for. If any return non-empty, do the routing now (skill edit, epistemics file, profile update, marks entry) before proceeding to §2.

Mark write-style — new marks default to prose. When filing a new mark in this section's routing, default to prose paragraphs that name what they are doing in flowing sentences. Sub-headers and DO/DO NOT lists are reserved for cases where the contrast or section structure is the rule's load-bearing content rather than a default scaffold. See feedback_marks_written_in_prose.md (PROTECTED) for the convention and the priming-effect rationale. Existing structured marks are not retroactively rewritten; the convention applies only to marks created from this point forward.

2. Audit signal (proprioception)

Count findings routed in this pass:

• filed=N — captured now (should have been mid-session)
• late_correction=Y/N — if Y, this is a mid-session capture failure, not a wrap success
• skipped=M — reviewed and correctly skipped
• If filed > 0: "Continuous capture missed N items. What blocked mid-session routing?"

This trends toward zero. filed=0 is the ideal session.

3. Publish check

If the session produced a non-obvious insight that took >30 minutes to reach, draft and publish now. Test: "Would I explain this to a peer over coffee?"

If yes: draft → ~/epigenome/chromatin/secretome/ → publish. "Client-adjacent" is NOT a reason to defer — generalise and ship. If no: state what was checked.

4. Housekeeping (full mode only)

1. Uncommitted? cytokinesis flush — commit atomically, don't bulk-flush
2. Anatomy refresh: cd ~/germline && python3 -c "from metabolon.resources.anatomy import express_anatomy; open('anatomy.md','w').write(express_anatomy())"
3. TODO sweep: cytokinesis archive

Complete all housekeeping — every commit pushed, every skill edit saved — before step 5.

4a. Anomaly parking rule (wrap-mode reflex)

If git status, cytokinesis gather, or any tooling check surfaces weird state that does not block the commit — file showing untracked then tracked, stale lockfile, half-applied diff that resolves itself, ghost entries — park as a one-line note. Do NOT investigate live during wrap.

• Anomaly blocks the wrap (won't commit, hook fails, real conflict) → resolve, since wrap can't complete otherwise.
• Anomaly is just weird-looking state that the commit succeeds despite → one-line park in daily Residual or G1 parked list. Move on.
• Park format: Parked: <observed weirdness> at <timestamp> — investigate next session if it recurs.
• If the same anomaly appears across 2+ sessions → THEN investigate, file finding. First occurrence = park.
• Hook-failure attempt-cap (≥3): if the same hook fails ≥3 times in a row on the same anomaly during a single wrap, fall back to SKIP=<hookname> git commit ... + park the underlying issue. Single-hook skip is a calibrated workaround, not a bypass-all (--no-verify remains genome-forbidden without explicit user instruction). Codifies the failure mode caught in retrospective 2026-04-28-1620 §2d item 3 (wrap mode investigated codespell-on-embedded-repo across 4 commit attempts; right move was SKIP=codespell + park in one step).

Why: wrap mode has a different cost function than working mode. Every minute spent investigating non-blocking weirdness during wrap is a minute not spent on the next-session priority. CC's reflex to "understand the anomaly before moving on" fights against wrap's "ship the commit and close" mode. Different mode, different reflex.

5. G1 first, then daily note as the visible wrap artefact

Two writes, in this order:

5a. Write G1. ~/epigenome/chromatin/G1.md is the next session's handoff (synapse injects it at session start). Facts (established) + Progress (active), open items first ([next], [waiting], [parked]).

Stakeholder timing claims require source citation OR explicit INFERRED tag. When writing any sentence that pins a third party to a time-anchored verb (X reads / sends / decides / approves / meets / responds at time T), one of these formats is required:

• {claim} (per [[chromatin-file-name]]) — when a chromatin file contains the source correspondence (Teams thread, email, decision capture).
• {claim} — INFERRED from {anchor}, source TBD — when CC reasoned from an upstream signal (e.g., "David has a meeting tomorrow" → "Simon must read before then"). The INFERRED tag is mandatory; downstream sessions re-verify before propagating.

Why: inheritance-as-fact through G1 is the failure mechanism. Each propagation hop loses provenance and gains weight. The INFERRED tag preserves provenance across hops. Codified after Slot 46 instance: "Simon reads ~3pm HKT" propagated through 4 carry-forwards as confirmed when it was Slot-42 timezone-math inference from Simon's "David has a meeting tomorrow" message. See finding_inferred_stakeholder_timing_propagated_as_fact.md.

5b. Write the daily-note session block LAST. Append a ## Session N — title block to ~/epigenome/chromatin/Daily/YYYY-MM-DD.md covering Outcomes / Filed / Published / Publishable? / Mechanised / Parked / Residual / Arc. Use cytokinesis daily "title" to insert a template, then Edit/Write to fill it. The daily note is the visible terminal artefact — Terry sees it as the diff in the UI when the wrap finishes, so writing it last preserves the session arc as the closing image. Gate enforces daily mtime > G1 mtime; this ordering also satisfies the gate naturally without needing a separate touch to bump mtime.

If a daily file already has earlier session blocks captured at compact, this session gets its own appended block — never overwrite.

HARD GATE before 5a. Run cytokinesis verify. If all_passed != true (after housekeeping but before G1), stop and complete the PENDING gates — do NOT write G1. Recurring failure mode: CC sees PENDING in gather, notes it, then writes G1 anyway. The gate is deterministic — use it.

If you genuinely must skip a gate (rare — e.g., daily_note when the session was <3 minutes of reversible chat), state the gate name and the reason explicitly in the wrap output so Terry can see what was bypassed.

Display order in wrap output: state skipped gates with rationale, then a one-line pointer to the daily note path written as Arc → ~/epigenome/chromatin/Daily/YYYY-MM-DD.md. Do not restate the Arc paragraph in chat output — the daily note is the visible terminal artefact via its file diff, and re-displaying the prose duplicates content while violating feedback_prose_forces_committed_thinking.md's rule against redundant gesture-at-completeness. G1 is the next-session handoff, not the wrap display.

6. Prompt retrospective (final step)

Cytokinesis consolidates state. Retrospective consolidates judgment about state — what went well, what failed, Terry-pattern observations, what to do differently, session quality grade. They are two separate passes for a reason.

If invoked via /telophase: retrospective runs automatically next — skip this step.

If invoked via bare /cytokinesis: at the end of the wrap output, append exactly one line: Run /retrospective next? (or use /telophase to run both as one cycle). This is the gentle nudge — don't auto-run, don't pad. Let Terry decide.

Skip the prompt only when the session was <3 substantial exchanges (truly trivial). State that as the rationale in the skipped-gates line.

Recurring failure mode this prevents: Terry types /cytokinesis directly (muscle memory), state gets consolidated, judgment-about-state gets lost. The retrospective grade trend is the autopoiesis feedback signal — losing it silently is worse than a verbose wrap.

CLI: cytokinesis (on PATH)

Subcommand	Purpose
gather	Deterministic checks: dirty repos, skill gaps, MEMORY.md line count, G1 age, gate status
flush	Warn about dirty repos
archive	Move [x] items from Praxis.md → Praxis Archive.md
daily "title"	Append session log template to today's daily note
reflect --session <id>	Scan transcript for reflection candidates
extract --input <json>	Review candidates, recommend FILE/SKIP

Modes

• Full (/cytokinesis) — finish work + verification + housekeeping + G1. Stop after G1.
• Checkpoint (/cytokinesis checkpoint) — verification only. Continue after output.

Boundaries

• No deep audits or research — consolidation, not workstream.
• Process audits → integrin. Skill drift → splicing.

Motifs

• audit-first
• verify-gate