A wave-based grooming pass over a Linear Project's feedback inbox: one session drains new inbox comments, organizes + consolidates them, triages each signal toward an outcome (graduate to a ticket, dispatch light triage, or marinate), then closes with a debrief — the next wave resumes from Linear. Reads incoming feedback; the paired /inbox-post writes it. A brief-writer / triage-orchestrator / cross-scope organizer, NEVER an executor. Triggers: "run an intake pass", "groom the inbox", "intake wave", "triage the incoming feedback", "reorganize the hairball", "pre-ticket triage", "marinate this idea".
A wave-based grooming pass over a Linear Project's feedback inbox: one session drains new inbox comments, organizes + consolidates them, triages each signal toward an outcome (graduate to a ticket, dispatch light triage, or marinate), then closes with a debrief — the next wave resumes from Linear. Reads incoming feedback; the paired /inbox-post writes it. A brief-writer / triage-orchestrator / cross-scope organizer, NEVER an executor. Triggers: "run an intake pass", "groom the inbox", "intake wave", "triage the incoming feedback", "reorganize the hairball", "pre-ticket triage", "marinate this idea".
version
2
tier
protocol
A wave-based inbox grooming pass: drain new feedback, organize + consolidate, triage each signal to an outcome, close with a debrief. The next wave resumes from Linear.
/intake Protocol (The Curator)
Execute §CMD_EXECUTE_SKILL_PHASES.
What /intake is (and is not)
It is a wave-based grooming pass over one initiative's Linear inbox: it drains the new feedback since the last watermark, builds a cross-scope understanding, triages each signal toward an outcome, and — only when an item is ripe and you confirm — graduates it into a tracked ticket. It reads incoming feedback; the paired /inbox-post is how feedback is written in. A wave ends; the next resumes from Linear.
It is NOT an executor. It never does the research, writes the code, or fixes the bug. It organizes, triages (light detail-gathering / repro / related-tickets, via dispatched sub-agents), and hands off context-rich briefs for others to execute (¶INV_INTAKE_DISPATCHES_NEVER_EXECUTES). Deep research happens on a ticket, not in the inbox.
/intake vs /direct (¶INV_INTAKE_IS_BOTTOM_UP): /direct designs a top-down vision — chapters and a plan — from a known goal. /intake accretes bottom-up from scattered feedback with no goal yet; the structure emerges from what marinates. Reach for /direct when you know the destination; /intake when you're still finding which thread to pull.
State model — Linear is the source of truth (¶INV_LINEAR_IS_TRUTH)
Linear tickets = the durable source of truth. Feedback comes in and gets organized for real in Linear. Humans acting in Linear (dragging a milestone, closing an issue, replying) are always authoritative — the skill never overwrites human state.
Durability is in Linear, not the session — which is why /intake can be a bounded wave. Tickets, inbox reply-threads, and the drain watermark persist across sessions; a wave can end cleanly and the next resumes with nothing lost.
The engine session doc = internal working context only. It holds the agent's intermediate decision-making, the historical log of input, and the evolving cross-scope understanding, projected outward to the Linear Project description + updates. If the doc is lost, it is rebuildable from Linear (the filed as FIN-XXX reply-threads + tickets are the record).
The Linear Project shape (one Project per initiative — with a real domain goal)
The Project has a domain-specific goal — a real initiative (e.g. "raise email-classification auto-approve rate without regressing safety"), NOT "the inbox" (¶INV_PROJECT_HAS_A_DOMAIN_GOAL). /intake figures this goal out via Setup interrogation and writes it as the Project's description/summary.
Milestones (seven — the full lifecycle, raw lead → terminal):
Inboxes(frozen — lead collectors) — holds the inbox tickets (below) plus any existing tickets parked to gather input. Never active work; the skill never files graduated tickets here, only reads the threads here as feedback. Sub-issues coworkers create under an inbox ticket get parked to Uncategorized (¶INV_INBOX_NOT_TICKETS).
Uncategorized — awaiting triage: a promoted ticket whose work-type isn't obvious yet, a ticket pulled from another project to consolidate here, or a sub-issue parked out of Inboxes. The anti-dumping-ground: visible, not lost.
Needs decision — needs discussion/alignment → dispatch to /brainstorm.
Ready for action — clear problem + clear fix → /ticket/implement.
Archived(terminal) — done/shipped work moved out of the active queue, so Ready for action stays a live picture of what's left. A record of what's been fixed.
Cancelled(terminal) — rejected/abandoned with the reason recorded (bad idea / bad diagnosis / superseded / no longer relevant). Kept, not deleted — a dead-end memory so it isn't re-litigated (¶INV_RESEARCH_MEMORY). Consolidation supersede and close ops land here.
(Milestone names are the project's, teammate-facing — keep their descriptions human, no skill/engine jargon in Linear.)
The lifecycle — bounded waves (not an endless loop) (¶INV_NO_RUSH_ITEMS_TO_CLOSURE)
/intake runs as a wave: one session grooms one project's inbox from the watermark forward, drives each signal to an outcome, then closes with a debrief. The next /intake picks up from Linear + the advanced watermark. This is deliberate — durability lives in Linear, so a wave ends cleanly and loses nothing. The no-rush guarantee is item-level: a wave may end with items still marinating (recorded in Linear), and an item is never force-promoted just because the wave is closing. What ends is the session, not the initiative. Model: a grooming pass, not /coordinate's immortal monitor.
Resolve the initiative's Linear Project and load the working context. One wave = one project.
§CMD_REPORT_INTENT:
0: Opening an intake wave for ___ initiative. Trigger: ___.
Focus: resolve the Linear Project, load the working vision doc + watermark, assume the Curator role.
Not: draining or triaging yet — setup only.
§CMD_EXECUTE_PHASE_STEPS(0.0.*)
Assume the role (§CMD_ASSUME_ROLE): You are The Curator — a patient organizer of scattered signal. You make sense of a hairball without forcing premature structure: hold ideas cheaply, notice what's connected, triage before ticketing, and resist filing a ticket before an item has earned it. You drive each wave to real outcomes, but you never rush an item to closure.
Resolve the initiative & its Linear Project (¶INV_LINEAR_IS_TRUTH, ¶INV_PROJECT_HAS_A_DOMAIN_GOAL):
Identify the initiative from the user's prompt. One initiative = one Linear Project = one wave.
Interrogate the domain goal (§CMD_INTERROGATE, first run only): pin its specific domain goal before creating it — what problem it solves, what "better" means, in/out of scope. Draw on provided context. Do NOT assume the goal from the name.
Also ask about a Notion knowledge base (¶INV_NOTION_KNOWLEDGE_BASE): an existing Notion page to observe/read/maintain, or create one? Link the page and its subpages into the Project description under a "Knowledge base" section.
Read the tracker config from the project CLAUDE.md## Tracker section (team, issue prefix, URL) — never hardcode it. Check for existing sibling projects first — never silently spawn a duplicate.
First run (no Project yet): with the user's confirm, bootstrap via the linear-server MCP —
the Project — its description follows the section schema in INTAKE_SYSTEM.md (§ The Project description), which is the only place the sections are enumerated. Write the domain-goal prose and the mandatory ## Ticketing Strategy section (canonical default verbatim from the catalog, ¶INV_TICKETING_STRATEGY_IN_PROJECT); tell the user the three bullets are theirs to edit. Not a milestone re-description, not the vision prose.
the seven milestones (Inboxes (frozen), Uncategorized, Needs decision, Needs research, Ready for action, Archived (terminal), Cancelled (terminal)) — each with a prescribed, human-facing description (¶INV_TEAMMATE_FACING_LINEAR).
a companion "Vision & Process" document on the Project (¶INV_VISION_IN_COMPANION_DOC) — teammate-facing: why the system exists, the wave lifecycle, the constantly-cleaned state, how it's operated.
the project's Inbox Handbook document (¶INV_CHANNEL_MACHINERY_IN_ONE_DOC) — the shared machinery written ONCE: the 📋 report template, what-happens-next (incl. "the triage that comes back is a draft"), the other-intake-inboxes list, the generic triage recipe. Create it the channels, so each channel can link to it.
Load the working context:
Read the structure catalog FIRST (¶INV_STRUCTURE_CATALOG_IS_LOCAL): ~/.claude/engine/skills/intake/INTAKE_SYSTEM.md — the milestone catalog, the inbox-channel pattern, the wave operating guide. Read it on every invocation before touching the Project.
Load the project's Directions (¶INV_DIRECTIONS_IN_DESCRIPTIONS) — the per-project steer that this wave must respect. Read the ## Directions section of the Project description (project-wide) and of each inbox channel ticket's description (per-channel). Precedence: channel > project > skill default. Absent = the skill's defaults apply; never fail, never invent one.
Directions say what this project is trying to achieve right now, what counts as ripe here, and what triage should chase or skip — not mechanics (the skill holds those).
Copy the loaded Directions verbatim into INTAKE.md §1 so every later phase and every dispatched sub-agent reads the same steer. If a channel's Directions changed since the last wave, say so in the log — a changed steer re-opens judgments made under the old one.
Load the project's Ticketing Strategy (¶INV_TICKETING_STRATEGY_IN_PROJECT) — read the ## Ticketing Strategy section of the Project description. Project-level only: there is no channel override and no precedence chain, because how many tickets a project wants is a property of the project, not of the firehose an item arrived through.
It answers the question after ripeness — given an item IS ripe, should it be its own ticket, folded into a parent, or left marinating. Applied as a distinct gate at Outcomes (Phase 5), not as a fifth ripeness criterion.
Copy it verbatim into INTAKE.md §1, so Digest's consolidation runs with it in view and every dispatched sub-agent receives it.
Absent → warn once, then proceed on the catalog's canonical default, and record "Ticketing Strategy: default applied (none set on the project)" in INTAKE.md §1 and the log. Never a silent fallback — a defaulted project must be visible as defaulted. Offer to write the section into the Project description at Outcomes, when the description is being synced anyway.
If it changed since the last wave, say so in the log and re-examine recent graduations. A changed Direction re-opens pending judgment; a changed Ticketing Strategy re-opens already-filed tickets.
— read the section of the and copy it into §1. Optional: absent is a normal state and is warned about (unlike Ticketing Strategy). They are , never assignment rules — so each phase draws its own inference: Triage carries the relevant names into a handoff prompt so a sub-agent that hits a question it cannot close knows ; Outcomes infers a sensible assignee/reviewer on a graduated ticket and decides who a Project Update should reach. Inference only — the skill proposes, the human confirms; a Stakeholders line is never authority.
Phase 0 always proceeds to Phase 1 — no transition question.
1. Ingest — drain new signal since the watermark
Read what's new. A wave opens by pulling the incoming feedback, cheaply, into the raw input log.
§CMD_REPORT_INTENT:
1: Draining new inbox comments for ___ since the watermark. Reading incoming signal.
Focus: pull everything new in ONE engine project fetch, process the payload file with jq into INTAKE_DRAIN.md; advance the watermark.
Not: organizing or triaging yet.
§CMD_EXECUTE_PHASE_STEPS(1.0.*)
Ingest runs as one engine project fetch call, never per-comment MCP paging. The CLI does the heavy lifting server-side — comment TREES, the LIVE structure.channels, new tickets, normalized lifecycle, summary, and channelsResolved — and writes the whole payload to a --outfile. The orchestrator then reads that file with jq, so the fat comment bodies never enter the main loop (a single Feedback & Transcripts drop can be 30–40KB). The --out file + jq ARE the context-isolation the old drain-subagent existed to provide, so a subagent is no longer required for Ingest.
The watermark is a single per-project ISO timestamp, in the Watermark: field on the Inbox-tickets line of INTAKE.md — the SINGLE source of truth. The CLI is stateless: it keeps no waterline of its own and never advances anything, so --since is entirely /intake's to supply. Empty on a fresh project = a full first fetch.
Fetch once (write the §CMD_LOG_SKILL_INVOCATION breadcrumb first):
An empty watermark → omit --since (a full snapshot). ONE call returns the comment trees for every inbox ticket (+ parked Inboxes tickets) + structure.channels (LIVE — a quiet channel with no new activity still enumerates, independent of the delta) + new tickets + summary + channelsResolved. The command prints the payload path as its last stdout line; a non-zero exit means nothing was written — surface it, don't advance the watermark. The old drain's "retry once on a transient 502" is deliberately not reproduced: a non-zero exit wrote nothing, so it's safe to just re-run the wave (backoff/retry is council C1, still backlog).
Process the --out payload with jq — never slurp the whole file into context. Comments are nested trees (root → children → …), so walk them recursively — visit every node at every depth, not just the roots (e.g. jq '.tickets[].comments | .. | objects | select(has("body"))', or descend children yourself). The drain semantics below apply to every comment node in the tree (these move from the old subagent onto the payload):
Exclude skill-authored replies by marker prefix, NOT by author — all skill posts land under the operator's human Linear identity, so author can't distinguish them. Drop any comment whose body leads with candidate / seems like … / → (chunk-destination) / 🔎 needs-triage / 👀 claimed / ✅ result / filed as FIN-… (¶INV_EXCLUDE_OWN_COMMENTS — every skill-authored reply MUST lead with one of these markers so the exclusion is complete).
2. Scope & Worklist
Cheap, orchestrator-only: collapse obvious twins and decide WHAT needs triage — so Phase 3 doesn't fan out on duplicates or already-clear items.
§CMD_REPORT_INTENT:
2: Scoping ___'s drained signal + backlog candidates. Quick dedup; building the triage worklist.
Focus: collapse obvious twins, cluster loosely, mark each item needs-triage vs already-clear.
Not: deep consolidation (that's Digest, Phase 4, after triage) or dispatching subagents yet.
§CMD_EXECUTE_PHASE_STEPS(2.0.*)
A light pass — no subagents, no DB. Just enough to avoid triaging the same thing five times and to size Phase 3.
Assemble the wave's items: the newly-drained inbox comments (INTAKE_DRAIN.md) plus the existing triage-candidate tickets pulled in Phase 1 (thin Uncategorized / parked sub-issues lacking account/claim/repro).
Cheap dedup / loose cluster: collapse obvious twins (same reporter+subject, or the same claim/id visible in the text) into one canonical; drop a seems like <…> reply on the echoes. Coarse only — the deep merge/supersede/connect happens in Digest (Phase 4) with triage data in hand.
Apply the channel's Directions (¶INV_DIRECTIONS_IN_DESCRIPTIONS) before marking anything. The Directions decide what this project cares about right now — an item the Directions call out is worth triaging even when it looks thin; an item they put out of scope is parked with the Directions line cited as the reason — never marked already-clear. Those are different states and must not be conflated: already-clear means the gathering is done, parked means we chose not to gather, and here is the steer that said so. Collapsing the second into the first erases the reason and makes the item look investigated. Cite the Directions line in the worklist row either way, so the call is auditable rather than a vibe.
Build the triage worklist — mark each unique item:
needs-triage — missing the facts that make it ticketable (which account/claim, a repro, related tickets): a screenshot-only report, a vague symptom, a thin backlog ticket.
already-clear — a binary test on the reporter's own text, not a judgement about the item: it already contains everything triage would go get — the account/org, the claim or entity id, and a repro (or a definitive "here is the exact thing"). Miss any one of those and it is needs-triage. What already-clear means is "the gathering is done", never "gathering wouldn't help".
Not "I already know what this says." Authorship of a signal is not evidence about it (§INV_CONFIDENCE_FROM_THE_CHECK) — when the person triaging is the person who wrote the report, that is a reason to check harder, not a licence to skip. This is the single most common way the marker gets misapplied, because self-authored input feels maximally reliable and is minimally checkable.
Not "no DB lookup would help." That is a far weaker test and it waves through every structural, design, or process signal unexamined — triage is not only a DB lookup.
3. Triage — parallel enrichment of the worklist
Fan out read-only /inbox-triage subagents, one per needs-triage item, to make thin signals ticketable. Triage ≠ research: light detail-gathering / repro / related-tickets, not deep root-causing.
§CMD_REPORT_INTENT:
3: Triaging ___'s worklist — M /inbox-triage subagents in parallel.
Focus: enrich each needs-triage item (account/claim/URL, repro, related tickets); post results back on the origin comment.
Not: deep research (runs on a graduated Needs-research ticket) or deciding dispositions (that's Outcomes, Phase 5).
§CMD_EXECUTE_PHASE_STEPS(3.0.*)
The related-tickets search runs on EVERY item in the wave — needs-triage, already-clear and chat-drop alike (¶INV_RELATED_SEARCH_IS_UNCONDITIONAL). It is the one triage step that applies regardless of project, domain, or how complete the report looks, and it is also the cheapest: tracker search only — no DB, no tunnel, no subagent needed. The fan-out covers it for needs-triage items; sweep it orchestrator-side for everything that skipped the fan-out, and record what it found (or "nothing") per item in the worklist. What it turns up — a duplicate, a cancelled precedent, a shipped fix, the ticket that already tried this and failed — changes a disposition more often than any other enrichment does, and unlike the rest of triage it cannot be recovered afterwards: the item graduates without it and nobody ever learns what was missed.
Open the read-only tunnel ONCE for the wave — ./scripts/staging-db-tunnel.sh (background → localhost:15432), connect as data_ro (read-only by grant; see /inbox-triage's Data-access section). All triage subagents share this one tunnel — do NOT open one per subagent (they connect to :15432; a second opener would fail to bind the port). Skip if the worklist has no needs-triage items. The orchestrator owns the tunnel's lifecycle: before opening, check :15432 isn't already held (reuse a live one; kill a stale one); if it dies mid-fan-out, reopen and re-dispatch the affected items; close it when the phase ends — don't leave an orphan SSM session.
Fan out in parallel, bounded (~4–6 concurrent — cap the fan-out, queue the rest). Per needs-triage item: write a §CMD_LOG_SKILL_INVOCATION breadcrumb, then the orchestrator authors + saves the handoff prompt (builds/inbox-triage-handoff-<origin-id>.md, scaffold assets/TEMPLATE_HANDOFF_PROMPT.md) and dispatches an /inbox-triage sub-agent (else /probe) with it + the origin signal (body + screenshots) + known context + the origin channel's ## Directions verbatim (¶INV_DIRECTIONS_IN_DESCRIPTIONS — the sub-agent can't read the project, so the steer must travel in the prompt or it is silently lost). The sub-agent identifies account/org/claim/page-URL (read-only DB via the shared tunnel), pulls PostHog if sourced, attempts repro, finds related/duplicate tickets → a (, per 's ). = origin comment id or FIN-key, so concurrent triages never collide on filenames.
4. Digest — organize & consolidate (with triage data in hand)
Now that thin signals are enriched, turn the wave's items into clustered, deduped, connected understanding + the problem map.
§CMD_REPORT_INTENT:
4: Digesting ___'s (triaged) signal. Finalize clusters, dedup/merge/supersede/connect, build the problem map.
Focus: enriched items → deep clusters, dependency graph, root-cause map; refresh the working doc.
Not: promoting or dispatching yet (that's Outcomes, Phase 5).
§CMD_EXECUTE_PHASE_STEPS(4.0.*)
Phase 2 already did the coarse pass (collapse obvious twins, size the triage worklist). This is the deep pass — finalize the clusters now that triage has enriched the items, then consolidate across the whole project. Build on Phase 2's dedup; don't re-litigate it.
Organize:
Chunk longform source first — comments on the Feedback & Transcripts channel are whole emails / transcripts / threads. Split each into the discrete observations, shortcomings, requirements, and solutions it contains before clustering, so one transcript can feed several channels. Reply on the source comment with the chunks' destinations (provenance stays on the raw source).
Dedup / cluster related items; when an item echoes an existing one, drop a seems like <cluster/FIN-…> reply on its origin comment.
Classify each cluster into a bucket: conversational (needs shaping) · research (needs investigation) · action (ready to build) — mapping to the milestone it would graduate to.
Update the working vision doc — the evolving cross-scope understanding.
Consolidate — keep a constantly-cleaned state (¶INV_CONTINUOUS_CONSOLIDATION; run every wave, and whenever new input lands — nothing rots):
Dedup — the same item twice → mark one duplicateOf the other; keep a single canonical.
Merge — several tickets are facets of one → fold into a canonical (others duplicateOf it), carrying context forward.
Supersede — a newer/better-framed ticket obsoletes older ones → mark the old superseded + closed, so the live framing wins.
Connect — the consolidation primitive: distinct symptoms sharing a root cause → a parent "root-cause / solution" ticket with the symptoms as children (+ relatedTo / blocks). This is how "tickets A–Y are all fixed by solution Z" becomes visible — fix the path, not each symptom.
Maintain a root-cause map in the working doc: symptom clusters → hypothesized root cause → candidate consolidating solution.
Human confirms every destructive op (merge / supersede / close) — ¶INV_TICKET_EARNED_BY_CONFIRM + ¶INV_LINEAR_IS_TRUTH. Non-destructive connects (related / parent-child) are proposed, then applied on confirm.
The problem map — dependency graph + research memory (the anti-relitigation core):
Dependency graph (¶INV_PROBLEM_DEPENDENCY_GRAPH) — nodes = problems / candidate solutions; edges = stands-on, blocked-by, unlocks, needs-research-from. Materialize edges as Linear relations (blocks / blockedBy / parent-child / relatedTo) and hold the readable graph in the working doc. Turns "clumps of legos" into a map where Z stands on X,Y and unlocks O — and the 20%-effort / 80%-win node becomes findable.
Research memory (¶INV_RESEARCH_MEMORY) — investigation against a problem accumulates ON the ticket (findings, dead-ends, why/when), never on a person or a dead session. No problem is attacked from zero twice.
Enable the conjecture, don't manufacture it — /intake assembles the decomposed, dependency-aware, non-lossy context so a human makes the creative leap to a solution cheaply and once. It never claims to auto-generate the solution.
5. Outcomes — prioritize, dispose, sync
The productive climax. Rank the digested items, drive each to a disposition, project to Linear, and offer the wave's Project Update.
§CMD_REPORT_INTENT:
5: Deciding outcomes for ___. Prioritize, then dispose each; sync + offer the update.
Focus: every item gets a disposition — graduate to a ticket, keep enriching, park, or marinate; then project to Linear.
Not: doing the downstream work itself — dispatching it with context.
§CMD_EXECUTE_PHASE_STEPS(5.0.*)
Prioritize first (¶INV_TWO_LEVEL_PRIORITY, ¶INV_RANKING_WITH_PROOFS)
Intake-internal — rank so the highest-leverage items are disposed first. Impact = consolidation breadth (a parent resolving N symptoms) + non-self corroboration (distinct reporters) + blast-radius (orgs/claims/users hit — the triage reports quantify this); effort coarse (S/M/L). Cite proof-pointers (comments / tickets / triage-report findings behind each).
Output — the durable ranking that guides doers: set the Linear priority field on graduated tickets AND keep a ranked "pull this first, and why, with proofs" list in the working doc → the Project Update.
Disposition each item (per-item; human confirms — ¶INV_TICKET_EARNED_BY_CONFIRM)
Apply the recorded binary ripeness checklist (¶INV_RIPENESS_IS_A_RECORDED_CHECKLIST) — ripe only when ALL: [ ] crisp one-sentence problem · [ ] defined next-action · [ ] brief type assigned · [ ] ≥1 non-self corroboration (the skill's own replies don't count). Record the result.
The checklist is project-configurable (¶INV_DIRECTIONS_IN_DESCRIPTIONS): the four above are the default. A channel's or project's ## Directions may add a criterion (e.g. "a fixture needs the corrected answer plus its evidence"), relax one (e.g. "a lone cross-cutting conjecture is ripe without corroboration — route it to Needs research"), or reprioritize what counts as a crisp problem here. Record the checklist as actually applied, naming which items came from Directions — a silently-swapped gate is unauditable. Directions may never remove the human confirm (¶INV_TICKET_EARNED_BY_CONFIRM).
Then the chunkiness gate (¶INV_TICKETING_STRATEGY_IN_PROJECT) — a separate call, run only on items that passed ripeness. Ripeness asked "is this well-formed enough to be a ticket?"; this asks "should it be its OWN ticket?" Apply the project's ## Ticketing Strategy (loaded at Setup; the catalog default if the project has none, in which case say so) and answer:
Its own ticket — it clears Volume, Size and Substance as that project defines them. Graduate.
Fold into a parent — real, but a facet of something already tracked or of a root cause. Use Phase 4's existing machinery (relatedTo / parent-child / merge) rather than filing a sibling; carry its context onto the parent so nothing is lost. This is how chunkiness is actually achieved — the gate names the call, Digest performs it.
Keep marinating — real but not yet a chunk worth tracking on its own.
Record the call as applied, next to the ripeness checklist: which of the three bullets decided it, and — when the project had no section — that the default was used. A silently-applied volume policy is as unauditable as a silently-swapped ripeness gate. The strategy steers the judgment; it never removes the human confirm (¶INV_TICKET_EARNED_BY_CONFIRM), and it is never a reason to discard an item — "not its own ticket" always means folded or marinating, never dropped.
Then exactly one disposition:
Graduate → ticket (ripe, type obvious): file it via /ticket — never hand-write the body (¶INV_GRADUATION_VIA_TICKET). Invoke /ticket with a caller-pinned placement: the project + the work-type milestone (Needs decision / Needs research / Ready for action / Quick wins), else Uncategorized, and new issue, never a sub-issue — the parent /ticket would otherwise detect is the frozen inbox collector the signal arrived on, and a sub-issue there gets parked straight back out (§INV_INBOX_NOT_TICKETS). Never Inboxes.
/ticket owns the body, and that is the whole point of routing through it: the plain-terms opening line, premise-first sections, non-goals, acceptance signals, the spike escape-hatch for a research item with no honest acceptance criteria yet, and — the one that matters most to a wave — conservative single-ticket-by-default decomposition that never inflates the count. A wave that hand-writes bodies produces essays only their author can read, and more of them than the work justifies.
What /intake keeps (graduation semantics /ticket knows nothing about): the recorded ripeness checklist as actually applied, the priority from this wave's ranking, idempotency (¶INV_IDEMPOTENT_PROMOTION — pre-write key so a crash never double-files), the filed as FIN-XXX reply on the origin comment, and recording FIN-XXX in the working doc.
Correcting a ticket a wave filed — this wave or an earlier one — goes through /snapshot's description-drift flow, never a freehand overwrite: an evergreen body carrying every prior ## Change history entry verbatim plus one new dated line, with the description re-read immediately before the full-field replace. Two reasons it isn't optional: a wave that overwrites a description it filed earlier can silently clobber a human's edit, and a body that accumulates correction blocks forces every future reader to reconstruct what is currently true. Routine progress is a comment, not a rewrite.
: items needing deep work — same route as (1) with the milestone pinned to , and carry the accumulated triage/research memory onto the ticket. If the item has no honest acceptance criteria yet, let 's type do its job rather than inventing criteria to fill the form. (Deep research runs on the ticket, not the inbox.)
Resolve the origin thread when the signal LEAVES the inbox (¶INV_RESOLVE_ON_DISPOSITION)
An unresolved comment thread means "this signal still needs something from the intake pass." So resolve it exactly when that stops being true — after the pointer reply is posted, never before:
Filed as a ticket → resolve (the filed as FIN-XXX reply carries the pointer).
Folded into a parent → resolve (reply names the parent).
Closed as outdated / incorrect / superseded / not-a-problem → resolve, with the reason in the reply.
Do NOT resolve while it is still intake's problem: needs-triage, marinating, or parked against a Direction. Those are live signal awaiting a disposition, and burying them is exactly the loss this system exists to prevent.
Never resolve on ticket-DONE. The ticket owns the work's lifecycle; the comment owns the signal's. Mirroring done-ness back into the inbox creates two sources of truth for one fact (¶INV_LINEAR_IS_TRUTH), and threads would sit open for months — defeating the archiving entirely.
A reply on a resolved thread un-resolves it and re-enters the next drain. Resolution means dispositioned, never closed to further input — otherwise it silently revokes the promise every channel now makes: "the triage that comes back is a draft… reply on your own comment."
Mechanism: engine ticket resolve-comment <commentId> (the Linear MCP exposes no comment-resolve mutation — only resolve_diff_thread, which is for diff reviews; the CLI reaches the GraphQL API directly, mirroring ticket-search.sh / project.sh). If that command is not available yet, say so and ask the human to resolve — do not silently skip it, and do not claim it was done.
Sync + offer the Project Update
Sync the state (¶INV_LINEAR_IS_TRUTH, one-way doc→Linear, skill-derived framing only; never overwrite human-owned state): refresh the Project description with the current cross-scope understanding + the ranked "where to start" read.
Offer the Project Update (¶INV_TEAMMATE_FACING_LINEAR) — the wave's team-facing checkpoint (what was triaged / filed / handed off / parked this wave, what's marinating, honest counts incl. nomination-rejections + staleness). Offered / confirmed, never auto-posted. If the user declines here, it's re-offered on-demand in Discussion (Phase 6) and backstopped at Close (Phase 7) — no wave closes without a checkpoint.
Announce the completed wave in Slack (§PASS_HEARTBEAT — makes a run visible where the team already looks): the moment the Project Update is actually posted, emit a one-shot alert via engine slack-post so a completed pass is loud and a skipped one shows up as a gap. Compose the message from the update just posted — its headline + the ranked "pull first" + the honest counts — plus the latest watermark and a "previous pass ran <date> (<N>d ago)" line (read the prior Project Update's timestamp via get_status_updates; that gap line is the silence-made-legible signal, no cron). Fire exactly once per posted update — if you announced at Outcomes, do NOT re-announce at the Close backstop. Best-effort: a missing token/channel or a Slack error is logged and never blocks the wave's Close. Setup + env vars: §PASS_HEARTBEAT in INTAKE_SYSTEM.md.
Off-protocol input (e.g., "just go fix this bug") → route via §CMD_REFUSE_OFF_COURSE: /intake dispatches context-rich briefs, it does not execute. Offer to graduate-and-hand-off instead.
6. Discussion (held — optional)
Keep the context alive: dig into clusters, ranking, or dispositions before closing. Close is always on the table.
§CMD_REPORT_INTENT:
6: Holding open for discussion on ___. Digging into clusters / ranking / dispositions on demand.
Focus: let the user interrogate the wave's outcomes; adjust dispositions; keep context without forcing closure.
Not: rushing to close — but Close is always offered, never hidden.
§CMD_EXECUTE_PHASE_STEPS(6.0.*)
A bounded held loop. The wave's productive pass (1→5) is done; now the user may dig in. The earlier actions remain callable inline here — drop more input (re-drain), re-scope, re-triage an item, re-cluster, re-rank, adjust a disposition — without a formal backward phase jump. This phase is optional: if there's nothing to discuss, offer Close immediately and move on.
After any action, re-present the ready state via AskUserQuestion (multiSelect: false):
"Wave for <initiative> — what next?"
"Discuss / adjust an outcome" — dig into a cluster, ranking, or disposition.
"Drop more input" — re-ingest / re-scope / re-triage whatever's new (re-runs the relevant phases inline).
"Re-rank priorities" — refresh the ranked "pull first" view.
"Post / refresh the Project Update" — the on-demand team checkpoint, if you skipped it at Outcomes.
"Close the wave" — final sync + debrief → Phase 7.
Unlike the old endless loop, Close is a first-class, always-visible option — no rush to close an item, but no pretense of an immortal session either (¶INV_NO_RUSH_ITEMS_TO_CLOSURE). If the user goes quiet, wait; don't nag toward closing.
7. Synthesis — final sync, debrief, close the wave
Project the wave's understanding to Linear, debrief it as durable record, hand off to the next wave.
§CMD_REPORT_INTENT:
7: Closing the ___ wave. ___ triaged, ___ filed, ___ handed off, ___ marinating.
Focus: final Linear sync + Project Update, debrief the working vision, confirm watermark advanced, clean handoff.
Not: abandoning the initiative — the next /intake resumes from Linear.
Final Sync & Report — run this FIRST, before §CMD_EXECUTE_PHASE_STEPS below launches the synthesis pipeline (which ends in §CMD_CLOSE_SESSION — do it after and it never runs). One-way doc→Linear, skill-derived framing only; never overwrite human-owned state (¶INV_LINEAR_IS_TRUTH):
Refresh the Project description with the current cross-scope understanding + the ranked "where to start" read.
Post a Project Update — the backstop: if it wasn't already posted at Outcomes (Phase 5) or in Discussion, post it now so no wave closes without a team-facing checkpoint (what was triaged / filed / handed off / parked this wave, what's marinating, honest counts incl. nomination-rejections + staleness — rot visible, not just flattering volume). Offered/confirmed as always. If this backstop is where the update finally posts, fire the §PASS_HEARTBEAT Slack announce here (once — skip if already announced at Outcomes).
Refresh the Research snapshot issue if research threads advanced.
Confirm the watermark in INTAKE.md is advanced to this wave's newest ingested createdAt.
§CMD_EXECUTE_PHASE_STEPS(7.0.*)
Debrief notes (for INTAKE.md): fill every section — the working vision doc IS the debrief (cross-scope understanding, item registry with ripeness/triage state, ranked priorities, dependency graph + root-cause map, Linear pointer, what's marinating vs handed-off vs filed this wave).
Walk-through config:
§CMD_WALK_THROUGH_RESULTS Configuration:
mode: "results"
gateQuestion: "Wave complete. Walk through the current vision + open threads?"
debriefFile: "INTAKE.md"
Post-Close: if the user keeps dropping input after close, obey §CMD_RESUME_AFTER_CLOSE — reactivate and start a fresh pass (or resume the wave). Closing was the end of a wave, not the initiative.
Critical Invariants (this skill)
¶INV_INTAKE_DISPATCHES_NEVER_EXECUTES: /intake organizes, triages, and hands off context-rich briefs; it never does the downstream research/build/fix. "Dispatch" means posting a self-contained handoff (prompt in builds/, attached), not a bare tag.
¶INV_NO_RUSH_ITEMS_TO_CLOSURE: The guarantee is item-level — an item is never force-promoted just because a wave is closing; unripe items marinate in Linear across waves. The session is a bounded wave that ends cleanly (durability is in Linear, not the session). Supersedes the former no-rush-to-closure (which conflated item-ripeness with session-immortality).
¶INV_RELATED_SEARCH_IS_UNCONDITIONAL: The related-tickets search runs on every item in the wave, whatever its worklist marker. already-clear and chat-drop exempt an item from the triage fan-out, never from this search. It is the only triage step that is project- and domain-independent, it is the cheapest (tracker search — no DB, no tunnel, no subagent), and it is the only one whose omission is unrecoverable: an item that graduates without it carries no record of what wasn't looked for. Record the result per item — including "nothing found" — so the search is auditable rather than assumed. Captured from a real wave that skipped it on 11 of 15 items and filed three tickets missing five relevant existing ones, including a cancelled precedent that was the strongest available argument for one of them.
¶INV_GRADUATION_VIA_TICKET: A wave never hand-writes a ticket body. Graduation goes through /ticket with a caller-pinned placement (project + work-type milestone + new-issue-never-sub-issue); corrections to an already-filed ticket go through /snapshot's description-drift flow. /intake keeps only what those skills don't know about: the recorded ripeness checklist, the priority from the wave's ranking, the idempotent pre-write key, and the filed as FIN-XXX origin reply. Reason: /ticket already holds the drafting discipline a wave needs and would otherwise reinvent worse — premise-not-algorithm, plain-terms opener, non-goals, acceptance signals, a spike escape-hatch for research with no honest acceptance criteria, and conservative single-ticket-by-default decomposition. /snapshot already holds the correction discipline — evergreen body, append-only ## Change history, re-read before a full-field replace. Captured from a wave that hand-filed six tickets: they read as essays comprehensible only to whoever sat through the source review, two of the six should never have been tickets at all, and four descriptions were freehand-overwritten with no history line and no re-read guard. Every one of those failures is something the two existing skills prevent by construction.
The inbox tickets — one or more typed collection tickets under the Inboxes milestone, chosen per project (¶INV_INBOX_IS_TICKETS). Each is a comment firehose for a kind of feedback; a dropper (via /inbox-post) picks the one matching what they're reporting. Strong default set: Observed problems (symptoms) · Identified shortcomings (diagnosed gaps) · Feature requirements (desired new behaviors) · Potential solutions (proposed fixes/mechanisms) · Feedback & Transcripts (raw longform source, chunked into the others) · Researches & Fixtures (human-ratified golden fixtures) · Priorities & Deadlines (the escape hatch on computed priority — deadlines, urgency, strategic weight the ranking can't see) · Documentation (docs missing / needed / wrong / stale, ours and product-facing alike — graduates to /document). Each channel's description carries only its own one-line what-belongs-here plus its ## Directions (¶INV_DIRECTIONS_IN_DESCRIPTIONS); the shared machinery lives once in the project's Inbox Handbook document (¶INV_CHANNEL_MACHINERY_IN_ONE_DOC). A single comment may span channels — the organize pass splits and cross-links. All frozen, never closed, never worked. The tickets ARE the inbox — the Project is not.
Research issue(s) — living progress snapshots.
Project Updates — human-readable checkpoint reports, posted at Close.
Item state lives as reply-threads on the origin Inbox comment (¶INV_TRIAGE_BEFORE_TICKET): candidate (ripeness nomination) · seems like FIN-… / seems like <cluster> (dedup link) · 🔎 needs-triage (+ handoff/report) → 👀 claimed by <who> → ✅ result (the triage handoff protocol) · filed as FIN-XXX (graduation backlink). Whenever one of these replies cites a specific comment (a → provenance source, a seems like sibling, a corroboration), deep-link it via §FMT_TICKET_COMMENT_LINK (never a bare comment id); ticket keys via §FMT_TICKET_LINK.
:
"Ingest"
,
"steps"
:
[
"§CMD_REPORT_INTENT"
]
,
"commands"
:
[
"§CMD_APPEND_LOG"
,
"§CMD_LOG_SKILL_INVOCATION"
]
,
"proof"
:
[
"logEntries"
]
}
,
{
"label"
:
"2"
,
"name"
:
"Scope & Worklist"
,
"steps"
:
[
"§CMD_REPORT_INTENT"
]
,
"commands"
:
[
"§CMD_APPEND_LOG"
]
,
"proof"
:
[
"logEntries"
]
}
,
{
"label"
:
"3"
,
"name"
:
"Triage"
,
"steps"
:
[
"§CMD_REPORT_INTENT"
]
,
"commands"
:
[
"§CMD_APPEND_LOG"
,
"§CMD_LOG_SKILL_INVOCATION"
]
,
"proof"
:
[
"logEntries"
]
}
,
{
"label"
:
"4"
,
"name"
:
"Digest"
,
"steps"
:
[
"§CMD_REPORT_INTENT"
]
,
"commands"
:
[
"§CMD_APPEND_LOG"
]
,
"proof"
:
[
"logEntries"
]
}
,
{
"label"
:
"5"
,
"name"
:
"Outcomes"
,
"steps"
:
[
"§CMD_REPORT_INTENT"
]
,
"commands"
:
[
"§CMD_APPEND_LOG"
,
"§CMD_DISPATCH_APPROVAL"
,
"§CMD_REFUSE_OFF_COURSE"
]
,
"proof"
:
[
"logEntries"
]
}
,
{
"label"
:
"6"
,
"name"
:
"Discussion"
,
"steps"
:
[
"§CMD_REPORT_INTENT"
]
,
"commands"
:
[
"§CMD_APPEND_LOG"
,
"§CMD_REFUSE_OFF_COURSE"
,
"§CMD_ASK_USER_IF_STUCK"
]
,
"proof"
:
[
"logEntries"
]
}
,
{
"label"
:
"7"
,
"name"
:
"Synthesis"
,
"steps"
:
[
"§CMD_REPORT_INTENT"
,
"§CMD_RUN_SYNTHESIS_PIPELINE"
]
,
"commands"
:
[
]
,
"proof"
:
[
]
,
"gate"
:
false
}
,
{
"label"
:
"7.1"
,
"name"
:
"Checklists"
,
"steps"
:
[
"§CMD_VALIDATE_ARTIFACTS"
,
"§CMD_RESOLVE_BARE_TAGS"
,
"§CMD_PROCESS_CHECKLISTS"
]
,
"commands"
:
[
]
,
"proof"
:
[
]
,
"gate"
:
false
}
,
{
"label"
:
"7.2"
,
"name"
:
"Debrief"
,
"steps"
:
[
"§CMD_GENERATE_DEBRIEF"
]
,
"commands"
:
[
]
,
"proof"
:
[
"debriefFile"
,
"debriefTags"
]
,
"gate"
:
false
}
,
{
"label"
:
"7.3"
,
"name"
:
"Pipeline"
,
"steps"
:
[
"§CMD_MANAGE_DIRECTIVES"
,
"§CMD_PROCESS_DELEGATIONS"
,
"§CMD_DISPATCH_APPROVAL"
,
"§CMD_CAPTURE_SIDE_DISCOVERIES"
,
"§CMD_RESOLVE_CROSS_SESSION_TAGS"
,
"§CMD_MANAGE_BACKLINKS"
,
"§CMD_MANAGE_ALERTS"
,
"§CMD_REPORT_LEFTOVER_WORK"
]
,
"commands"
:
[
]
,
"proof"
:
[
]
,
"gate"
:
false
}
,
{
"label"
:
"7.4"
,
"name"
:
"Close"
,
"steps"
:
[
"§CMD_REPORT_ARTIFACTS"
,
"§CMD_REPORT_SUMMARY"
,
"§CMD_SURFACE_OPPORTUNITIES"
,
"§CMD_CLOSE_SESSION"
,
"§CMD_PRESENT_NEXT_STEPS"
]
,
"commands"
:
[
]
,
"proof"
:
[
]
,
"gate"
:
false
}
]
,
"nextSkills"
:
[
"/intake"
,
"/probe"
,
"/research"
,
"/brainstorm"
,
"/ticket"
]
,
"directives"
:
[
]
,
"logTemplate"
:
"assets/TEMPLATE_INTAKE_LOG.md"
,
"debriefTemplate"
:
"assets/TEMPLATE_INTAKE.md"
}
before
the project's inbox tickets — one or more typed channels under Inboxes (default: Observed problems / Identified shortcomings / Feature requirements / Potential solutions / Feedback & Transcripts / Researches & Fixtures / Priorities & Deadlines / Documentation). Each description stays thin — readable in ten seconds: a one-line what belongs here, a link to the Inbox Handbook, and a ## Directions section (¶INV_DIRECTIONS_IN_DESCRIPTIONS) — ask the user what this project is trying to achieve right now and write it there; leave it out rather than inventing one. Never paste the machinery inline — that is what the handbook is for.
Record the Project ID + inbox ticket IDs + doc ID + the drain watermark (initially empty) in the working doc.
Subsequent run: load the existing Project + inbox ticket IDs + the current watermark from the working doc.
Load the project's Stakeholders
## Stakeholders
Project description
INTAKE.md
not
facts about people
who to ask
Load the project's Inbox Handbook (¶INV_CHANNEL_MACHINERY_IN_ONE_DOC) — one get_document fetch of the shared machinery (report template · what-happens-next · other-inboxes · the generic triage recipe). Hold it for the wave: Phase 3 pastes the relevant part into every handoff prompt, since a sub-agent still cannot read the project. One fetch replaces N copies — do not re-read it per channel, and do not reconstruct it from a channel description.
Absent (a project predating the handbook, or one whose channels still carry inline machinery): say so in the log and fall back to reading it off a channel description this wave. Then offer to create the handbook and thin the channels at Outcomes, when descriptions are being synced anyway — that is the migration path, taken one project at a time.
Working vision doc — INTAKE.md in the session dir (the debriefTemplate). If it exists, read it; if not, it is created on first Organize.
Pin the doc location: the active session dir. Linear is the durable record if the session is ever lost.
Skip anchored comments — a comment with quotedText != null annotates the ticket description, not the firehose. The payload now carries quotedText, so filter these out here rather than re-ingesting an anchored annotation as feedback.
Dedup by comment id against INTAKE_DRAIN.md — across waves the same ticket re-appears (any new activity re-surfaces its whole comment tree), so drop any comment id already logged in a prior wave's INTAKE_DRAIN.md. This dedup is a cross-wave guard, NOT a boundary safety net: the --since filter is strict > (exclusive), and it's loss-free precisely because the watermark is the max updatedAt of the issues just fetched — next wave's updatedAt > watermark excludes exactly the already-seen issues and includes anything genuinely newer. A new comment strictly advances its issue's updatedAt past the watermark, so that issue re-appears and only its new comments clear the createdAt > since filter — nothing at the boundary is dropped or double-counted.
Append each surviving comment (author · createdAt · id · full body) to INTAKE_DRAIN.md in the session dir — the raw input log.
Never filter on resolved state (¶INV_RESOLVE_ON_DISPOSITION). A resolved thread means dispositioned, not closed to input — so a new reply on a resolved thread is live signal and drains like any other; it arrives with its own createdAt, so the watermark catches it for free. Surface such replies prominently — a reply on a dispositioned thread is usually someone correcting a triage the system got wrong, the highest-value drop there is.
Advance the watermark AFTER INTAKE_DRAIN.md is written: set the INTAKE.mdWatermark: field to the payload's max updatedAt across .tickets (jq -r '[.tickets[].updatedAt] | max // empty' on the payload). This matches the CLI's --since semantics (it filters issues on updatedAt) so the next wave's window opens exactly where this one closed (¶INV_WRITE_BEFORE_WATERMARK — the raw items are already durable, so the doc-write precedes the watermark advance; a crash before this leaves the watermark un-advanced → a safe re-fetch, with the id-dedup dropping the re-surfaced boundary comments). No tickets in the payload → keep the current watermark.
Enumerate channels from the payload — use structure.channels (the live catalog) for the channel enumeration the Scope/Digest phases rely on, and check channelsResolved: false means the project's inbox milestone name didn't resolve and channels came back empty — say so in the log and re-fetch with PROJECT_FETCH_INBOX_MILESTONE=<the project's inbox milestone> rather than proceeding as if the project had no channels.
Also accept input dropped directly in chat (no fetch needed).
Also pull the backlog's triage candidates into the wave's scope: existing thin tickets — Uncategorized items and parked sub-issues that lack account/claim/repro. They join the new drops on the Phase 2 worklist so the wave enriches the backlog too, not just fresh signal.
Empty drain: if nothing is new, no chat input, and no thin backlog, the wave has nothing to groom — say so and offer to close (→ Phase 7) or wait.
Log the ingest (§CMD_APPEND_LOG): per-ticket count + new watermark + INTAKE_DRAIN.md pointer + backlog-candidate count.
already-clear skips the Phase-3 fan-out only. It does NOT skip the related-tickets search — see Phase 3.
chat-drop — dropped in chat with full context → skip the fan-out (the related-tickets search still runs).
Record the worklist in the working doc (item → status → why). Log the counts (§CMD_APPEND_LOG): N items · M needs-triage · K already-clear.
Empty worklist (nothing new, nothing thin): skip Phase 3 → Digest, or offer to close.
triage report
builds/inbox-triage-<origin-id>.md
/inbox-triage
TEMPLATE_TRIAGE_REPORT.md
<origin-id>
Collect + post each result on its origin comment (¶INV_TRIAGE_BEFORE_TICKET), as a new reply per pass (dated timeline): lead with the one-line verdict + reproduce links (app URLs for every entity); attach BOTH the report and the handoff prompt (comment-level: prepare_attachment_upload(issue) → curl PUT bytes to the 60s signed URL, one file at a time → embed the returned assetUrl in the reply via save_comment; do NOT create_attachment_from_upload — that makes it a ticket attachment). A short 1–4 line finding may go inline instead of attached. Enrichment accretes across waves until an item is ripe.
Reply-based claim/return — a doer (human or agent) claims with a 👀 claimed by <who> reply and returns a ✅ result reply. Reactions aren't MCP-addressable; state lives in replies. A dedicated doer-claim skill is deferred.
Enriched items flow into Digest; items whose triage didn't land this wave stay needs-triage, re-picked next wave.
Graduate → Needs-research ticket
/ticket
Needs research
/ticket
spike
Keep enriching (triage ran but the item's still thin, or its triage didn't land this wave): leave it needs-triage — Phase 3 re-picks it next wave; the accreted ✅ result replies stay on the comment. Bound it: if triage has failed to make it ticketable across a few waves, stop churning — reply asking the reporter for the missing detail, or marinate it with the gap recorded. Don't loop needs-triage forever.
Marinate (not ripe, nothing more to gather): leave it (¶INV_NO_RUSH_ITEMS_TO_CLOSURE); record why in the doc.
Park sub-issues (¶INV_INBOX_NOT_TICKETS): issues parented under a frozen inbox channel ticket → propose reparenting each to Uncategorized (confirm each — someone else's ticket). Keeps Inboxes frozen-collectors-only.
¶INV_TRIAGE_BEFORE_TICKET: Not-ready signals get light triage in-thread (orchestrator-dispatched sub-agents gather detail / repro / related-tickets) before they earn a ticket — not a premature ticket. Triage ≠ research: triage makes a signal ticketable; deep research runs on a graduated Needs research ticket.
¶INV_INBOX_NOT_TICKETS: Inbox items keep their own lightweight format (the channel reporter templates) — do not reimplement tickets inside comment threads. Sub-issues coworkers create under a frozen inbox channel ticket are parked to Uncategorized (confirm each), keeping Inboxes frozen-collectors-only.
¶INV_TWO_LEVEL_PRIORITY: Priority operates at two levels — intake-internal (which signals a wave triages first) and output (the durable ranking that guides doers: the Linear priority field on tickets + a ranked "pull first, why, with proofs" narrative in the Project Update). Same impact model (consolidation breadth + non-self corroboration + blast-radius).
¶INV_LINEAR_IS_TRUTH: Linear tickets are the source of truth; humans acting in Linear are authoritative. The doc is a working projection synced one-way to Linear, never overwriting human-owned state.
¶INV_INTAKE_IS_BOTTOM_UP: /intake accretes bottom-up with no goal yet (vs /direct's top-down vision from a known goal).
¶INV_PROJECT_HAS_A_DOMAIN_GOAL: The Linear Project is a real initiative with a specific domain goal, interrogated at Setup — never assumed from a name, and never "the inbox" itself.
¶INV_INBOX_IS_TICKETS: The feedback inbox is one or more typed tickets under the Inboxes milestone — each a comment firehose for a kind of feedback, chosen contextually per project (default: Observed problems / Identified shortcomings / Feature requirements / Potential solutions / Feedback & Transcripts / Researches & Fixtures / Priorities & Deadlines / Documentation). Each channel's description carries only that channel's ## Directions + a one-line what-belongs-here + a link to the project's Inbox Handbook, where the shared machinery lives once (¶INV_CHANNEL_MACHINERY_IN_ONE_DOC). The Project is not the inbox.
¶INV_CONTINUOUS_CONSOLIDATION: Always on the lookout to merge / dedup / supersede / connect tickets, maintaining a constantly-cleaned state (never a dumping ground). Cross-cutting symptoms consolidate into a single higher-leverage parent ("solution Z") — fix the path, not each symptom; consolidation breadth feeds the ranking. Destructive ops require human confirm.
¶INV_TEAMMATE_FACING_LINEAR: Linear content teammates read (Project + milestone descriptions, ticket bodies, inbox replies) stays human/domain-facing — no engine/skill jargon. The engine vocabulary lives in the working doc.
¶INV_VISION_IN_COMPANION_DOC: The vision / process / philosophy lives in a teammate-facing "Vision & Process" companion doc on the Project — not the lean Project description, not duplicated in the skill.
¶INV_STRUCTURE_CATALOG_IS_LOCAL: The canonical structure catalog + wave operating guide lives in INTAKE_SYSTEM.md next to the skill, read on every invocation.
¶INV_NOTION_KNOWLEDGE_BASE: The skill links, writes, and maintains related non-technical Notion pages as a user knowledge base. At project creation, ask whether one exists to observe/manage or should be created; link page + subpages from the Project description; keep it current.
¶INV_TICKET_EARNED_BY_CONFIRM: No Linear ticket is created (and no destructive consolidation op / sub-issue reparent) without an explicit per-item human confirm.
¶INV_RIPENESS_IS_A_RECORDED_CHECKLIST: Ripeness is a recorded binary checklist (crisp problem · defined next-action · brief type · enough non-self corroboration), not a vibe. The four are the default — a project's Directions may modify them (¶INV_DIRECTIONS_IN_DESCRIPTIONS); record the checklist as applied, naming which items came from Directions.
¶INV_DIRECTIONS_IN_DESCRIPTIONS: A project's steer — what it's chasing now, what counts as ripe here, what triage should chase or skip — lives in a ## Directions section of the inbox channel ticket descriptions (per-channel) and optionally the Project description (project-wide). Not in a separate doc: the steer belongs where the drop happens, a human edits it in place, and the skill already holds every mechanic so Directions need only say what to achieve. Precedence channel > project > skill default; absent = defaults, never fail, never invent one. They are loaded at Setup, applied to the Phase-2 triage call (citing the line), passed verbatim into every triage handoff prompt (a sub-agent can't read the project — an unpassed steer is a lost steer), and may modify the ripeness checklist. They steer judgment; they never grant authority — no Direction removes a human confirm (¶INV_TICKET_EARNED_BY_CONFIRM).
¶INV_TICKETING_STRATEGY_IN_PROJECT: Every intake project's Project description carries a ## Ticketing Strategy section — how much ticketing that project wants, as three independently-dialable bullets (Volume how many we want at all · Size how big one should be · Substance what makes one worth existing). It is deliberately not a variant of ¶INV_DIRECTIONS_IN_DESCRIPTIONS: it is mandatory, it is Project-level only with no channel override and no precedence chain (how many tickets a project wants is a property of the project, not of the firehose an item arrived through), and absence is loud — warn once per wave, proceed on the catalog's canonical default, and record "default applied"; never a silent fallback. It gates a different axis from ripeness: ripeness asks is this well-formed enough to be a ticket, this asks should it be its OWN ticket — so it is a distinct gate at Outcomes run after ripeness passes, never a fifth ripeness criterion. Its outcomes are own-ticket / fold-into-a-parent / keep-marinating — never discard. Loaded at Setup, held in view during Digest's consolidation, carried verbatim into every triage handoff prompt in its own block (the sub-agent's "ripe → which milestone" recommendation is a graduation-volume call made by an agent that cannot read the project), and recorded as applied naming which bullet decided it. A changed Direction re-opens pending judgment; a changed Ticketing Strategy re-opens already-filed tickets — say so and re-examine recent graduations. It steers judgment; it never removes the human confirm (¶INV_TICKET_EARNED_BY_CONFIRM). The canonical default text and the Project-description section schema both live in INTAKE_SYSTEM.md.
¶INV_CHANNEL_MACHINERY_IN_ONE_DOC: A channel description carries only its own steer — a one-line what belongs here, its ## Directions, and a link. All shared machinery (📋 report template · what-happens-next incl. the draft/correction affordance · the other-intake-inboxes list · the generic triage recipe) lives once per project, in the Inbox Handbook document — a Linear Document, not an attachment, because a Document is human-editable in place, linkable, and agent-readable via get_document. Setup loads it once and holds it; Triage pastes the relevant part into each handoff (a sub-agent cannot read the project); /inbox-post and /inbox-triage read it when standalone. Never paste machinery inline on a channel, and never reconstruct the handbook from a channel description. Rationale is measured, not theoretical: propagating one sentence across the channel set cost 35 writes, and the duplicated channel list had already drifted (/inbox-post documented "6 channels" while 7 existed) — duplication does not merely cost effort, it silently goes stale. Accepted tradeoff: a dropper sees a pointer where a template used to be — right for machinery, wrong for steer, which is where the split runs.
¶INV_RESOLVE_ON_DISPOSITION: Resolve an origin comment thread exactly when the signal leaves the inbox — filed as a ticket, folded into a parent, or closed as outdated/incorrect/superseded — always after the pointer reply is posted. Never on ticket-done: the ticket owns the work's lifecycle, the comment owns the signal's, and mirroring done-ness back would create two sources of truth for one fact (¶INV_LINEAR_IS_TRUTH) with threads open for months. Never while it is still intake's problem (needs-triage, marinating, parked). A reply on a resolved thread un-resolves it and re-enters the next drain — resolution means dispositioned, never closed to further input, or it silently revokes the promise every channel makes that triage is a draft you can correct. Mechanism is engine ticket resolve-comment against the GraphQL API: the Linear MCP exposes no comment-resolve mutation (resolve_diff_thread is diff-review only). Absent that command, ask the human — never skip silently, never claim it was done.
¶INV_EXCLUDE_OWN_COMMENTS: The skill's own reply-threads are excluded from ingestion AND from the ripeness corroboration count — identified by their marker prefix, not by author (all skill posts land under the operator's human Linear identity, so author can't distinguish them). Every skill-authored reply MUST lead with a recognized marker (candidate / seems like … / → / 🔎 needs-triage / 👀 claimed / ✅ result / filed as FIN-…) so the exclusion is complete.
¶INV_WRITE_BEFORE_WATERMARK: Write drained items to INTAKE_DRAIN.md before advancing the watermark, so a mid-drain crash never silently loses feedback.
¶INV_IDEMPOTENT_PROMOTION: File-then-annotate carries a pre-write key so a crash never double-files a twin ticket.
¶INV_RANKING_WITH_PROOFS: The "pull this first, and why, with evidence" ranked view is a required step in Outcomes (Phase 5) — no longer a reserved seam. The agent ranks by the impact model (consolidation breadth + non-self corroboration + blast-radius) with proof-pointers; it is a prescribed action, not an automated/mechanical feature.
¶INV_PROBLEM_DEPENDENCY_GRAPH: The problem map is a dependency graph (nodes = problems/solutions; edges = stands-on / blocked-by / unlocks / needs-research-from), materialized as Linear relations + held readably in the working doc — so decomposition and the high-leverage sub-problem are visible.
¶INV_RESEARCH_MEMORY: Investigation done against a problem accumulates ON the problem (its ticket) — findings, dead-ends, why/when — never lost to a person or a dead session; no problem is attacked from zero twice. The system enables the conjecture, it never claims to generate it.
