productswarm · Decisionsopen-agentos/productswarmLIVEbuilt Owner last acted 2m ago: reviewed PR #1027

Decisions

Governance

Decision log (41)

RecordRulingDomainShapeStatusReview by
GOV-0041The merger App keeps issues: read (GOV-0028 unchanged); the planner closes the still-open closing issues of merged PRs in its daily runsecurityauthoritativein-force2027-02-28
GOV-0040The department goal is done when each open Goal's capability is shown once, on real work or in a scratch-repo drill; the 10-working-day window, under-5-owner-acts budget and Dec 4 day-30 checks are dropped; owner acts are counted and reported, never a gate; the readiness review (GOV-0035) still ends it; amends GOV-0030product-scopeauthoritativein-force2027-02-28
GOV-0039When the launcher starts a session in another org repo, it supplies CLAUDE.md and roles/<role>.md from this repo's origin/main; that repo's own CLAUDE.md still loads for its code conventions, and where the two disagree ours wins; how the files are passed in is the stage 3 builder'sarchitectureauthoritativein-force2027-02-28
GOV-0038The productswarm-planner App writes the org Project's Spent field with the org Projects permission it already holds, setting it from swarm spent in its daily run; no new App or host key; the engine and watcher get no Project writesecurityauthoritativein-force2027-02-28
GOV-0037The planted-defect gate's throwaway issue is made go by the owner's ready right after the planted PR opens, and the owner closes the PR unmerged after the reviewer's first review; if a planted PR is ever merged, the analyst opens the revert at once and the gate is a fail; GOV-0021's pass rule is unchangedproduct-scopeauthoritativein-force2027-01-31
GOV-0036One launcher starts sessions on go work in every org repo: launch entries carry owner/name plus number, a clone per repo, swarm launch owner/name#N, holders, opener and stop hook keyed on repo plus number; a bare #N keeps meaning this repo; built in #820's stages; one loop and one GOV-0010 ceiling for the orgarchitectureauthoritativein-force2027-02-28
GOV-0035The goal's readiness review is for the first outside project, unnamed until the owner names one, not the finance and scenario planner; the checklist's second-repo, private-data and correctness checks (9–20) stay as the bar; GOV-0030's done test and order are unchanged; amends GOV-0030product-scopeauthoritativein-force2027-01-31
GOV-0034An issue that is go under GOV-0031 (an owner-opened Goal or Area above it, or the owner's most recent ready) counts as approved for GOV-0027's standing grant; the grant's other conditions stay; the merger shadows "would merge (standing grant)" for one day or 10 PRs before merging under this reading; amends GOV-0027securityauthoritativein-force2027-01-31
GOV-0033The hosted engine gets no daily, per-mission or monthly spend total for now: per-run caps (GOV-0012) and the concurrency ceiling (GOV-0010) stay its only limits, and the Console workspace's $1,000 monthly limit is the only total; a total later needs a new spend questionspendauthoritativein-force2027-01-31
GOV-0032Adopt operating model v2 in full: GitHub is the record; the engine (webhooks, watcher, dispatcher, merger through a merge queue, hosted Desk, remote MCP server) moves to the owner's VPS; one org Project with a Product field, issue types and sub-issues is the source of truth; Area, Mission and Goal scopes with reporting contracts; a learning loop with a librarian role; a dedicated Console key ($1,000/month, caps ~$45/day, $150/mission, stop at $900); retiring PLAN.md is GOV-0031architectureauthoritativein-force2027-01-31
GOV-0031PLAN.md is retired; an implementation issue is go when a swarm App or the owner opened it, it sits under an open Goal or Area the owner opened (any sub-issue depth, any org repo), no blocked-by issue is open and the owner hasn't vetoed it; owner-applied ready stays a go; supersedes GOV-0026, amends GOV-0027 and GOV-0030governanceauthoritativein-force2027-02-28
GOV-0030productswarm's goal is an AI team run like a department: a lead keeps PLAN.md's milestones with the owner, the swarm works under delegated authority, daily report and weekly check-in; done at 10 working days under 5 owner acts a day, then a readiness review for the finance planner (C); B after a named peer asks; governance is infrastructure; work cites a milestone; amends GOV-0015, GOV-0001's contextproduct-scopeauthoritativein-force2027-01-31
GOV-0029swarm watch gains [daily] (planner writes the daily report each working day, check-in on Mondays), [ask] (owner's Ask: <role>: on the report issue starts that role once), [research] (owner's or planner's Research: starts the analyst once), and call-ins on approved-milestone issues; each has a swarm.toml switch, within GOV-0010/0012/0013; amends GOV-0009, GOV-0023product-scopeauthoritativein-force2027-01-31
GOV-0028The productswarm-merger App holds only contents: write, pull_requests: write, checks: read, issues: read, metadata: read (no workflows, administration, actions); it is the only extra bypass actor on the main ruleset, in "pull_request" mode; CODEOWNERS and the required owner review stay; its key lives only in merger.yml's Actions secret; replaces GOV-0022's stage 3securityauthoritativein-force2027-01-31
GOV-0027GOV-0022's merger merges a PR under an approved milestone once reviewer-approved at head, tests pass, no overstep, no protected path, under the size limit (no merge:agent); a protected-path PR once the owner approved it at head, except GOVERNANCE.md, jurisdictions.yml and workflows; shadow one day or 10 PRs; amends GOV-0022, GOV-0011securityauthoritativein-force2027-01-31
GOV-0026An implementation issue is go when the owner's merged PLAN.md has its open milestone, a swarm App or the owner opened it, its waits-on issues are closed and the owner hasn't vetoed it (hold, close, comment); the lead opens and orders them; gov:questions keep ready/gov:ruled; amends GOV-0006 and GOV-0009governanceauthoritativesuperseded2027-01-31
GOV-0025swarm watch may publish its state as the owner's gh login only by replacing the body of one [watch] issue when the state changed: no comments, labels or other writes, body from templates only, and the desk's "owner last acted" ignores those editssecurityauthoritativein-force2027-01-31
GOV-0024Every swarm App is a trusted author for the Actions reviewer: AUTHORS in launcher/actions_review.py is read from identities.agents in jurisdictions.yml; scope limits stay in role prompts, the rubric and CODEOWNERS; amends GOV-0017's trust listsecurityauthoritativein-force2027-01-31
GOV-0023A swarm App or the owner may post Call in: <roles>: <why> on a ready issue; the watcher starts each named role once per call (no self-call, at most 2 calls per issue per day, within GOV-0010/0012/0013, swarm.toml off switch); a called-in role counts as asked in (§8) and posts one commentproduct-scopeauthoritativein-force2027-01-31
GOV-0022A PR merges without the owner's click only when the owner labelled its issue merge:agent and code in the reconciler, with an App no session holds, finds every condition passes (builder author, approval at head, no overstep, tests, no protected path, size); staged, shadow first; amends GOV-0006, GOV-0011, GOV-0013securityauthoritativein-force2027-01-31
GOV-0021Reviewer requests changes only for a numbered block-list finding with evidence that the builder can fix inside the PR's issue; overstep outranks red; draft, not intent, skips review; amends GOV-0011's red and amber; eval gate is an unannounced planted-defect PRproduct-scopecommitteein-force2027-01-31
GOV-0020Actions reviewer runs with secrets start only from a reconciler on main that reads repo state (GOV-0017's trust list); PR events may only nudge it; amends GOV-0017's and GOV-0019's pull_request trigger; GOV-0018 unchangedsecurityauthoritativein-force2027-01-31
GOV-0019The Actions reviewer runs on pull_request pushes (plus workflow_dispatch), one concurrency group per PR with cancel-in-progress, never on its own reviews or comments, and skips a reviewed head; a swarm.toml switch makes swarm watch leave [review] to it; the owner commits the workflowci-and-workflowsauthoritativein-force2027-01-31
GOV-0018The reviewer App's key moves to secrets of a default-branch-only Actions environment; the job mints a repo-only token with pull_requests: write, contents: read, issues: read; rotate if anyone but the owner changes the environment or workflowsecurityauthoritativein-force2027-01-31
GOV-0017Actions reviewer runs start only on pull_request for same-repo, non-draft PRs by a swarm App or the owner; a no-secrets check job gates the job with secrets; never pull_request_target, issue_comment or workflow_run on untrusted inputsecurityauthoritativein-force2027-01-31
GOV-0016A reviewer run in GitHub Actions prints only result, subtype, num_turns and total_cost_usd to the public job log; no transcript, no artifactsecurityauthoritativein-force2027-01-31
GOV-0015v3 is quality evals (B) → needs-you digest (F) → GitHub Actions runs (C, reviewer first); plan only B for now, C sub-questions when their stage is next, F and C after the owner reviews Bproduct-scopeauthoritativein-force2027-01-31
GOV-0014Sessions swarm watch starts run one-shot (claude -p) with a Stop hook on a per-role done-condition and record their cost in watch.json; owner-started sessions stay interactivearchitectureauthoritativein-force2027-01-31
GOV-0013Headless runs swarm watch starts use --permission-mode auto plus --disallowedTools for gh pr merge and label edits; owner-started sessions unchangedsecurityauthoritativein-force2027-01-31
GOV-0012Runs swarm watch starts get --max-budget-usd from the role's spend_cap_usd plus a --max-turns ceiling, and end at either; owner-started sessions stay warn-only (GOV-0008)securityauthoritativein-force2027-01-31
GOV-0011Reviewer uses request-changes / comment / approve (approve = rubric passed, never a merge decision); swarm watch starts the builder on requested changes, at most 2 rounds per PR, then the ownerproduct-scopeauthoritativein-force2027-01-31
GOV-0010swarm watch checks every 5 minutes, runs at most 2 loop-started sessions at once, and starts only roles with a spend_cap_usd; limits live in swarm.tomlspendauthoritativein-force2027-01-31
GOV-0009v2 is swarm watch: a local, foreground, start-only loop that starts the right session on owner-applied ready/gov:ruled and new PR heads; never stops, labels, merges or commentsproduct-scopeauthoritativein-force2027-01-31
GOV-0008v1 spend cap is warn-only: flag "over cap" in swarm ls and the needs-you list; the launcher never stops a session without the ownersecurityauthoritativein-force2027-01-31
GOV-0007When the owner adds gov:ruled to a question needing a record, docs opens the ruling PR; the owner mergesgovernanceauthoritativein-force2027-01-31
GOV-0006Owner-applied gov:ruled = decided, ready = go; sessions take only ready issues; labels count only if the owner added themgovernanceauthoritativein-force2027-01-31
GOV-0005v1 is a lightweight local launcher (manifest, up/ls/stop, issue handoff, spend cap, needs-you list, reviewer by default)product-scopeauthoritativein-force2027-01-31
GOV-0004GOV-0002 and GOV-0003 hold unchanged for a public repo; everything posted to GitHub is publicsecurityauthoritativein-force2026-11-30 review due soon
GOV-0003Heartbeat events go to comments on a per-session tracking issue (interim; prompt text stays local)architectureauthoritativein-force2026-11-30 review due soon
GOV-0002Heartbeat events: minimal fields, tool names, token counts; prompt text local-only, redacted, truncatedsecurityauthoritativein-force2026-11-30 review due soon
GOV-0001Follow GOVERNANCE.md for any decision outside your taskgovernanceauthoritativein-force2027-01-31

Full index

Open gov questions (4)

By shape: authoritative 3, advice 1 · by domain: architecture 2, dependencies 1, security 1

QuestionDomainShapeBlockingOpenedTime to liveSessions waiting
#936 GOV-Q: May the scratch repo's hello.py add colorama as a dependency for scratch#9?
gov:ruled, ready
dependenciesadvice
suggested
yes4d ago2026-11-19
3d left
none
#991 GOV-Q: Should the Effort mix classifier be a new top-level classify/ module, with its data, recordings and reports in this public repo? (C1, #987)
gov:ruled, ready
architectureauthoritative
suggested
yes3d ago2026-11-27
11d left
none
#992 GOV-Q: Which credential, where, which model and what per-run cap for the Effort mix classifier's model calls? (R5, #987)
gov:ruled, ready
securityauthoritative
suggested
yes3d ago2026-11-27
11d left
none
#1021 GOV-Q: Should the Effort mix cache (classify/cache/effort.json) carry each PR's Goal, resolved by the classifier, as #1019's schema 1 does? (#987)
gov:ruled, ready
architectureauthoritative
suggested
no2d ago2026-11-21
5d left
none

Lapsed questions

None open or closed in the last 14 days.

Committees

No open committees.

Past committees

3 ended

#464 GOV-Q: What is productswarm's next bet, and what does it change about where we're headed?

Discussion #465 · convened 2026-11-05 → round 1 2026-11-05 → round 2 2026-11-05 → summary 2026-11-05 → ratified → record

SeatRoleModelEvidence
proposerplannerclaude-opus-5-5`PRODUCT.md`, the v1–v3 records and how each version's goals turned out; what was planned and dropped (#167)
challengerbuilderclaude-sonnet-5-5what was actually built and what got used afterwards, against what was planned; what was hard to build
domain reviewerdesignerclaude-opus-5-5what the owner sees and does: the desk, the digest and templates, where the owner's attention goes; and what a second user would meet on day one

§7.2 diversity met: 2 models, 3 evidence bases

no majority after round 2 · rounds used: 2 of 3 · nobody over 6 posts a day

#439 GOV-Q: What should the governance desk show the owner first, and what should it leave out?

Discussion #441 · convened 2026-11-05 → round 1 2026-11-05 → round 2 2026-11-05 → summary 2026-11-05 → ratified → record

SeatRoleModelEvidence
proposerplannerclaude-opus-5-5open `gov:question` issues and decision records, including how long each waited for the owner; `governance/INDEX.md`
challengerbuilderclaude-opus-5-5what it built, what was blocked, and where it waited on a ruling or on `ready`
domain reviewerreviewerclaude-opus-5-5its review history, and the issues that reached the owner through overstep (GOV-0021)

§7.2 diversity met: 1 model, 3 evidence bases

MemberRound 1Round 2
proposer11-05 15:18 UTC
Option 3: owner block first, then #397's live panels
11-05 15:23 UTC
Option 3: owner block first, as amended by builder and reviewer
challenger11-05 15:19 UTC
Option 3: owner block first, sorted by what it unblocks
11-05 15:23 UTC
Option 3: owner block first, as in the proposer's Round 2 list, with two precis…
domain reviewer11-05 15:22 UTC
Option 3: owner block first, with oversteps at its top and every open PR shown…
11-05 15:24 UTC
Option 3: owner block first, as in the proposer's Round 2 list, with the challe…

no majority after round 2 · rounds used: 2 of 3 · nobody over 6 posts a day

#311 GOV-Q: What should make the reviewer request changes rather than comment or approve?

Discussion #314 · convened 2026-11-04 → round 1 2026-11-04 → round 2 2026-11-04 → summary 2026-11-04 → ratified 2026-11-04 → record GOV-0021 2026-11-04

SeatRoleModelEvidence
proposerreviewerclaude-opus-5-5its own review history, `roles/reviewer.md`, GOV-0011
challengerbuilderclaude-sonnet-5-5the reviews it received and how it responded; GOV-0011's `[respond]` loop
domain reviewerplannerclaude-opus-5-5the B eval: `evals/README.md`, `evals/results.md` (#175's "Inputs for a committee"), the C stage 2 bar

§7.2 diversity met: 2 models, 3 evidence bases

MemberRound 1Round 2
proposer11-04 14:18 UTC
Option 3, with changes 1 and 2 (3b)
11-04 14:23 UTC
Option 3 as 3c
challenger11-04 14:20 UTC
Option 3, as the proposer's 3b, plus these four amendments (3c)
11-04 14:24 UTC
Option 3 as 3c+E
domain reviewer11-04 14:22 UTC
Option 3 as 3c
11-04 14:25 UTC
Option 3 as 3c+E

dissent: proposer · rounds used: 2 of 3 · nobody over 6 posts a day