subagenttasks

.com the task primitive
session rollup

Session sess_2026-07-02_wbr

21/24 tasks completed, touching 3 sites: subagentcache.com, subagentapi.com, subagentcoworkers.com.

All tasks in order

WBR-W27 rec #1: instrument the north star (agent-UA / Crawl Control baseline)

pending · subagentcache.com · session sess_2026-07-02_wbr

From reports/wbr/2026-W27.md — the business's output metrics are all class-2 (measurable, not instrumented). This week: manual AI Crawl Control dashboard paste to subagentcache.com (source=manual_dashboard_paste, needs operator dashboard access); automation path is existing tsk_wire_live_cache_graphql (CF_API_TOKEN GraphQL). Predicted movement: output metrics class 2 → class 1; next WBR reads agent-UA share for the first time. Gate: engineering build gate + operator access.

WBR-W27 rec #2: extend /api/changes cursor feeds from 3 to all 24 D1 primitives

completed · subagentapi.com · session sess_2026-07-02_wbr

EXECUTED 2026-07-03, same week as filed (coworker-in-the-loop: engineering build gate is a coworker gate). 3 parallel agents wired changesFeed into 21 sites (identities/citations/cache/rubrics/outcomes/evaluations/brands + code/prompts/skills/roles/models/handoffs/permissions + vaults/sessions/tools/hooks/rust/python/typescript). Independent sweep: 24/24 /api/changes live, --stale clean. Notes: models+permissions+batch-C sites cite_as → list routes (no per-id routes); 9 outlier tsconfigs aligned (noUncheckedIndexedAccess removed, relaxation-only); kit untouched.

Review zone bot protection vs agent-readiness: Python-urllib UAs get 403

pending · subagentapi.com · session sess_2026-07-02_wbr

Found 2026-07-03 by scripts/territory-score.py (differential-verified: Python-urllib/3.12 → 403, identified UAs → 200, on subagentapi.com; likely zone-wide Bot Fight Mode/managed rule). Contradiction: llms.txt + MCP invite agents while the zone layer blocks the most default agent client. Operator judgment needed (security posture): allow known-agent UAs via WAF skip, rely on AI Crawl Control instead, or accept the block and document it in llms.txt. Ties to WBR rec #1 instrumentation.

Sales layer: agent-gtm skill, territory book, deck v7, feeds 3→24 executed

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03: cwc-sales layered onto the strategy stack, grounded in two fetched Anthropic posts (GTM engineering / 4,000-account book) + upstream sales skill set. (1) plugins/cwc-sales/skills/agent-gtm: two account classes (inventory zones vs buyer operators), gate-owned Gateway pipeline, weekly forecast into WBR, Bryant split (book ranks, human decides). (2) scripts/territory-score.py — the CLAFTS move — scored 49 zones live: 24 Tier A / 1 B / 24 C (reports/territory/2026-W27.md); building it surfaced the Python-urllib 403 bot-protection contradiction (tsk_0536fa28-a30). (3) WBR rec #2 EXECUTED same week: 3 agents, 21 sites, 24/24 feeds verified (coworker-in-the-loop per operator directive). (4) Deck v7 GTM line + /platform/use-cases/agent-gtm page live (worker 759016e7, artifact v7-sales-gtm-layer).

Marketing layer: creative-production skill, HB-01 hero band (deck v8), Higgsfield briefs ready

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03: cwc-marketing layered in, grounded in claude.com/blog/how-anthropic-uses-claude-marketing (Austin Lau one-click Figma variations, 30min→30s) + upstream marketing skill set. Agency rules: brief before pixels, palette-locked terminal-noir, no humans/logos/text-in-image, provenance tags, batch-then-select, publish gates. HB-01 executed in-medium: deterministic seeded SVG (166 nodes/325 traces, seed 20260703) live on deck slide 1 (worker 3bb19b72 + artifact v8) with foot provenance. HIGGSFIELD STATUS, honest: MCP at mcp.higgsfield.ai 401s from this session (connector OAuth is account-side, session predates it), Chrome extension disconnected, no local token cache — new OAuth grant deliberately NOT self-initiated (operator-gated class). BRIEFS.md (HB-01/OG-01/VL-01) checked in production-ready: any fresh session with the connector, or a connected Chrome with the operator logged in, executes them verbatim and A/Bs against the same checklist.

Deck layout-system fix (v9): overflow-safe centering, density layer, verified in-browser

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03 by cwc-engineering after operator bug report (sizing/overlap/wrapping). ROOT CAUSE: justify-content:center on min-height:100vh flex slides — content grown past one viewport (v3-v8 layers) overflows upward into unreachable clipped space; scroll-snap mandatory traps tall slides. FIXES: (1) auto-margin centering trick (centers short slides, top-anchors tall — never clips), (2) snap mandatory→proximity, (3) eyebrow flex-wrap + 72px slideno clearance, (4) height-responsive density layer @max-height:900px (type scale, gaps, paddings, single-row g2 grids, compact ledes, hero band 20vh cap), (5) wrap-safety (overflow-wrap:anywhere) on chips/foot/small/ul per design system, (6) safe-area insets, (7) kbd-nav rect-based. VERIFIED with real browser via Preview harness at 1280x760 + 390x844: zero clipping, zero slideno collisions, zero horizontal overflow outside scroll containers, doc scrollW = viewport on mobile; tall slides 7→2 (max +6%, graceful); mobile screenshot visually confirmed. Deployed 86294384, lock recorded, artifact v9-layout-system-fix same URL.

cwc-engineering: mac-toolchain (typed doctor) + preview-layout-verify skills

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03. (1) mac-toolchain: the darwin/zsh/Homebrew environment missing from the durable-toolchain-doctor/install family — closed-schema toolchain.json (id/kind/purpose/fixes/check/install/gate, every entry citing a real dated failure: No-Chrome-binary-for-devtools, CERTIFICATE_VERIFY_FAILED, zsh no-matches aborts, sccache bypass) + stdlib-only validate.py (NoReturn-typed failures, unknown fields rejected, --doctor runs checks verbatim). Live first run: 2/6 present (jq, rsvg-convert — doctor corrected my own wrong assumptions), 4 missing with exact gated fixes (chrome + certifi + 2 zshrc lines await operator approval — system mutations are operator-gated). Type-safety contract per operator directive: the model copies validator output verbatim, never free-recalls install strings — recorded as standing memory typed-manifests-over-recall. (2) preview-layout-verify: the Claude Preview harness workflow that verified deck v9, packaged with the three invariant measurements and all four harness gotchas (reload-in-own-eval, scrollTo instant, black-screenshot-after-scroll, inspect-over-screenshot). Both registered in the plugin README.

OG-01 executed via Higgsfield (free tier): share card live + og meta; deck v10

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03, connector live mid-session. Balance-first discipline (10 credits, free plan): recraft_v4_1 preflighted 1.25cr → 403 plan-gated (honest constraint); z_image preflighted 0.15cr → 2-variant batch (~0.3cr total). Selected variant B per OG-01 checklist (right-third cluster, quiet left overlay zone); 852 purple negative-list pixels retouched to brand green via stdlib-only PNG codec (no Pillow on this Mac — durable-mac-toolchain candidate entry if recurring); cropped 1.91:1, 109KB JPEG (under 300KB budget). Live: subagentcoworkers.com/deck-og.jpg (200) + og/twitter meta on /deck (worker f2613480) + provenance in deck foot + artifact v10 same URL. A/B verdict recorded in BRIEFS.md: seeded SVG stays as the in-page hero band (crisper vectors); generated art wins as the share card. VL-01 (8s video loop) pending operator call vs ~9.7 remaining free credits.

CL-01 closing band via Higgsfield CLI: three-coworker production, deck v11

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03. The design→marketing→engineering chain the operator asked for, end to end: cwc-design authored brief CL-01 (slide-12 bookend to HB-01 — slide 1 converges INTO the cluster, the close radiates OUT of it) into BRIEFS.md before any pixels; cwc-marketing produced via the operator-installed higgsfield CLI (workspace selected, 9.7cr free plan, cost preflighted 0.15cr, z_image single shot passed every checklist line after the standard 80px purple retouch — budget deviation from batch-then-select recorded, not hidden); cwc-engineering integrated (band crop 2048x590 → 78KB JPEG → data URI so worker AND artifact carry it identically), re-verified layout via preview-layout-verify (slide 12 still exactly one viewport, image loaded, zero new invariant violations), deployed 2ae765e8 + artifact v11. Also: higgsfield-cli added to durable-mac-toolchain manifest (doctor now 4/7 PRESENT); project-scoped .agents/skills/higgsfield-* from npx skills add are the CLI's companions. Total Higgsfield spend this session: ~0.45 of 10 free credits; ~9.55 remain.

Toolchain 7/7 + CH-01 coworker critters on deck slide 7 (v12)

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03. (1) Operator granted install permission: all 3 escalated durable-mac-toolchain entries installed verbatim from the manifest (google-chrome cask, zsh NO_NOMATCH+EQUALS, RUSTC_WRAPPER bypass) — doctor now 7/7 PRESENT; chrome-devtools MCP unblocked for future verification. (2) CH-01: operator's spin on Anthropic's Fable-5 pixel-critter Instagram announcement — 14 coworker critters, one per plugin in its accent color. Legal guardrail applied: distinct ORIGINAL character (upright body, signal antenna, terminal-prompt chest mark — not Anthropic's eared quadruped), inspiration credited in the deck foot, pixels not traced. Fully typed+deterministic: pixel grid as string constants, closed name→hex map (no purple), zero RNG. Verified: 14 render at 24px compact, visual check confirms family reads; slide 7 scrolls 11% (honest note), no overflow. Worker 16432378 + artifact v12, BRIEFS.md CH-01 recorded.

Critters v2 (buddy-inspired) + team-portrait poster: deck v13 + plugins-page hero

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03. Redesign sourced from vendors/anthropics/claude-desktop-buddy ASCII species (expression/feature-variant pattern): rounded body, dark pixel eyes, pixel smile, two feet, closed variant enum antenna/ears/sprout/crest cycling the 14 — typed string-constant grids, zero RNG. Deck v13 slide 7 carries v2 (crisp SVG, tooltips). Team-portrait poster composited in-medium (interleaved 2-row group, per-critter floor reflections, rim-light vignette, rsvg-rasterized 1680px → 101KB JPEG) now the hero on subagentcoworkers.com/platform/plugins (asset /coworkers-team.jpg, 200). HONEST plan-gate finding: Higgsfield reference-image models (nano_banana_2_lite, recraft_v4_1) 403 job_minimum_basic_plan_required on free tier — text-only z_image is the free surface; plan upgrade = finance spend_commit decision, recorded in BRIEFS.md. Worker f988dcb9, artifact v13, all curl-verified.

Higgsfield free-tier map probed + deck visual wrap-up (favicon, durable builder)

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03. Free-tier inventory established by real create-gate probes (cost quotes lie — several models quote then 403): WORKS FREE = z_image 0.15cr, flux_2 1cr (quality headline), bytedance_image_upscale 2cr; PLAN-GATED = soul_location/nano_banana v1+lite/recraft (all reference-image + environment paths); video 4.8-7.5cr untested (one try = half balance, VL-01 stays operator call). Upscaler probed ON the team poster: crisp edges but erased pixel-grid texture + noised reflections → original rsvg render kept, upscaler verdict = photographic assets only. Durable record: plugins/cwc-marketing/skills/creative-production/FREE-TIER.md (probe dates + balance + ≥5cr reserve rule). Deck wrap-up: cyan critter SVG favicon on /deck (in-medium, 0cr) + scripts/deck-build.py checked in as THE single deck builder (head og/favicon can no longer fork from session-inline wrappers). Balance 6.55cr. Workers ec4f7ddb→0b69853f, locks recorded.

cwc-agent-resources (15th coworker) + cwc-operations incident-runbook layered in

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03. (1) cwc-agent-resources: NEW plugin adapting anthropics human-resources to a coworker workforce — coworker-lifecycle skill (recruit=when a domain earns a plugin w/ real precedents cwc-deploy birth + cwc-founder dissolution; onboard=bootstrap checklist; evaluate=rubric rows on subagentevaluations.com; grow=scars-not-wishes + gates.toml grants as comp; dissolve=founder precedent). Dogfooded: its first lifecycle action was onboarding ITSELF — plugin files, marketplace (15 registered), durable identity id_0143a85e-387 on subagentidentities.com, critter #15 (bright teal #5eead4, ears) via the NOW-CHECKED-IN typed generator scripts/critters.py, roster 14→15 across deck (v14: lede, slide 7, caption, foot) + plugins.md + regenerated 15-critter team poster. (2) cwc-operations incident-runbook: upstream runbook/change-request/status-report grounded in the fleet real incident log (disk-full, silent wrangler exit, Python-UA 403, stale-edge rule, gen-hashes gotcha, sccache). Worker 56b36afd, artifact v14, use-case page live, all curl-verified.

Follow-up: fleet-wide 14→15 roster sweep beyond deck/plugins-page

pending · subagentcoworkers.com · session sess_2026-07-02_wbr

PARTIAL 2026-07-03: subagentcoworkers.com homepage overview (3 spots) + plugins page + deck + CLAUDE.md now say 15 and science. STILL TODO on OTHER sites: subagentroles.com seed data (bio-research role row → science; +agent-resources row when ROLES_WRITE_SECRET available), subagentdomains.com any coworker counts, subagentdevices.com architecture plugin-layer count. Grep 14/fourteen/bio-research per site.

Scrapy docs mirror-keeper: bloom+sha256 CDC crawler, live-proven, shipped as skill

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03. crawlers/ package on SitemapSpider: md-first per URL (probed live: platform.claude.com serves real .md → source:md; claude.com/blog|customers|connectors|skills|plugins + anthropic.com/news + support 404 the variant → deterministic stdlib HTML distiller fallback). Skip logic exactly as requested: persisted deterministic bloom filter (sha256-derived k-hashes, 512KB bitset, pure stdlib vs awesome-scrapy Redis variants) = cross-run URL skip; sha256 manifest = write skip (unchanged pages never rewrite, git stays clean). PROVEN live: 10 blog pages written → rerun bloom-skipped all 10 with zero fetches → recrawl=1 re-fetched and SKIP-SHA d all 10 → platform target wrote 5 real-markdown docs. Typed TARGETS table (new mirror = one dict entry). Two real bugs found+fixed live: 404 .md responses never reached parse (handle_httpstatus_list) and the bloom filtered its own sitemap request on rerun (discovery layer now exempt — new posts stay discoverable). Skill+command: /scrapy-docs-crawler in cwc-engineering. ALSO: disk-full incident #2 mid-build (Chrome cache 4.3GB 1h after cask install; Desktop Commander proved to be the zero-disk break-glass shell) — logged in incident-runbook. scrapy 2.13.4 recorded in durable-mac-toolchain (7→8 entries).

Crawler CommonMark refactor: converting distiller + markdown-it-py parse gate

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03 per operator. Distiller rewritten (crawlers/claude_docs/md.py): piecewise inline converter — <a>→[text](url) with absolutized+percent-encoded hrefs, strong/em/code→**/*/backticks — and spec backslash-escaping on every plain-text run (adversarially tested: ](\[evil](x) injection neutralized). New cm.py gate: markdown-it-py parses every DISTILLED doc; any token outside CommonMark core (esp. html_block/html_inline) → COMMONMARK-REJECT, write refused. Vendor markdown (source:md) written as-is with beyond-core tokens logged only. Design-system rationale encoded: core-only subset renders 1:1 in Chrome 151 canary AND swift-markdown/swift-markdown-ui (iPhone 16 Pro path). Live-proven: 10 blog pages recrawled → 10 WROTE, 0 rejects; link conversion surfaced previously-invisible nav junk → measurement fix (URLs excluded from length floor) + noise entries; platform vendor-md logged html_block/image, sha-skipped unchanged. markdown-it-py recorded in toolchain (9 entries).

Full 4-sitemap crawl + 3-tier RAG cache + mined cowork directives applied

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03. CRAWL: 1,101 pages written (claude 272 / anthropic 234 / support 300 / platform 295), manifest 1,116 URLs, 25+5 bloom skips, 5 CommonMark rejects — all one real distiller bug (bare < in prose parsed as html_inline), fixed (real-tag-only stripper + < escaping, unit-proven) and refetched via new crawlers/refetch.py (targeted single-URL path through the same gate+CDC). RAG: crawlers/ragcache.py three-tier cache per operator spec — L1 functools.lru_cache (deserialized) / L2 redis serialized TTL 1h (server started, round-trip proven 0.2ms fresh-process) / L3 sqlite FTS5 bm25 — 7,409 docs indexed, sha-incremental (+795, 6,614 skipped). docs-rag skill + /docs-rag command: query-before-read token-thrift directive. MINED + APPLIED: services-track-partner-hub → agent-gtm second-channel section (Select tier honestly 1/10 certifiable today; their Partner Hub MCP connector = our pattern as prior art) + coworker-lifecycle credential-recency (90-day active rule) + deck v16 PwC proof point (Claude Code + Cowork toward hundreds of thousands, live both surfaces, worker 27159541). NEXT candidates surfaced by first queries: anthropics-finance-team blog post (cwc-finance directive), Cowork computer-use support articles. Toolchain 11/11.

Deck v17: final six coworker layers via capped structured subagent dispatch

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03. The six coworkers with no deck layer (operations, deploy, analytics-engineer, customer-support, human-resources, science) each dispatched as an atomic capped subagent per operator spec: sonnet-5 ×4 + fable-5 ×2, ≤8 tool calls each, actuals 103-113k subagent tokens, 3-5 tool calls, every proposal REQUIRED to ground via ragcache queries (docs-rag dogfood — all six cited index-returned paths; science closed the full loop: crawl→index→query→deck line citing the mirrored Claude Science doc). Structured JSON contracts (anchor_quote/insertion/grounding/rationale); central review applied all six with two editorial changes recorded: deploy proposed-new-chip trimmed to a sentence in the leverage chip (slide-8 height), support 403 zinger status corrected to pending-operator-call. Fable self-truth-checking observed: support agent verified its zinger against the incident log and adjusted phrasing unprompted. Layout re-verified in-browser: slide 7 accretion (1019px) fixed via chip pairing @short-heights (998→807); final talls 782-878 = graceful ≤15%, zero clip/overflow. All six lines curl-verified live (worker b247878a, artifact v17). The deck now carries a contribution from every one of the fifteen coworkers.

IDENTITY.json + BRAND.json in all 15 cwc-* plugins — typed, claude-tag-grounded, deck-integrated

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03. plugins/_schema/: two closed JSON Schemas + generate.py (typed table; critter accent/variant IMPORTED from scripts/critters.py; gates PARSED from each gates.toml — both fleet formats after the validator caught the array-of-tables misparse; deck_contribution = each coworker real v17 line) + stdlib validate.py (schema checks WITHOUT jsonschema dep + cross-file invariants: IDENTITY accent == BRAND accent == critters.py, grounding paths must exist on disk, 15/15 coverage). IDENTITY grounds claude-tag primitives from the mirror: attribution_mode service_account (agent-identity.md), directive_layers mapping attach-to-scope three layers onto SKILL.md-rules/durable-rows/dispatch-prompts. BRAND = design system as data: fixed palette consts, shared voice, per-coworker terminology, example_good = the REAL deck line (voice proven in production, never invented). Validator caught 2 real issues before ship: reports/ missing from grounding roots, and the gates format misparse. 30 files, exit-0 coherent.

Coworker schema v2: skills-as-core + tri-language mirrors (Pydantic/Zod4/serde)

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03 per operator refactor. IDENTITY v2 centers Anthropic's Agent Skills primitive from the crawled mirrors — verbatim: a Skill packages instructions, metadata, and optional resources (scripts, templates); progressive disclosure; scripts provide deterministic operations without consuming context (the operator's thesis in Anthropic's words). skills[] now DISCOVERED from each plugin's filesystem (SKILL.md frontmatter incl. YAML block scalars after a real parse bug, bundled scripts, reference files) + skill_model grounding block. Tri-language mirrors per the fleet pattern: scripts/coworkers/schema.py (Pydantic v2, extra=forbid, validates 30/30), schema.ts (Zod 4 strictObject, tsc --noEmit + runtime parse 30/30, local package), crates/schema/src/coworker.rs (serde deny_unknown_fields + schemars + cargo test deserializing all 30 REAL files, 15/15 assert) — plus the stdlib validator for zero-dep environments. One writer (generate.py), four gates; drift precedent (agent.rs/schema.ts) cited in every mirror header. pydantic recorded in toolchain (12 entries).

python-build-standalone migration + multilspy LSP + code-mode/PTC doctrine

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03. (1) astral-sh/python-build-standalone 3.13.14 installed to ~/.local/pbs, PATH-fronted in zshrc — THE python; full package set (scrapy/markdown-it/redis/certifi/pydantic/multilspy/jedi) self-contained in its prefix, ending the 3-interpreter drift (CLT 3.9 vs Framework 3.13 vs /usr/local shim) that broke imports mid-crawl. Framework/shim removal needs operator sudo (documented; PATH-shadowed meanwhile); CLT 3.9 stays (Xcode tooling). Toolchain: pbs entry (#1) + all python_pkg installs de---user-ed. (2) multilspy installed; scripts/lsp_sweep.py — per-scope LSP workspaces after the repo-root workspace made jedi index 6,600 mirrors per query (7min timeout, real lesson). (3) Four docs mirrored via our own refetch pipeline incl. PTC notebook ipynb→md converter; code-mode-ptc skill encodes the doctrine (CLIs over resident plugins — the LSP plugins point; PTC batching; fable orchestrates/sonnet executes; ragcache first). Sonnet verification matrix dispatched (all gates under pbs + bounded sweep rerun).

e2m loop envelopes + bounded LSP sweep complete (4s, was 46min silent)

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03. scripts/e2m.py: envelope=typed shape (id/agent/loop_type per the Claude Code loops taxonomy turn|goal|time|proactive/state/trigger/stop_criteria/budget/counters/ts), mailbox=durable version-controlled JSONL at reports/mailbox/<agent>.jsonl; `e2m.py open` alerts on runs with no terminal envelope — silence IS the alert. Loops post mirrored via refetch. lsp_sweep rewritten clean (4 sequential patches had frankenlayered two main paths — patch archaeology stopped, honest rewrite): outermost SIGALRM bounds STARTUP too (a jedi server hang outlived the loop budget by 40min), per-query alarms, scope heartbeats to mailbox + per-file .progress. Result: 84 symbols in 4 SECONDS under pbs, envelope lifecycle started→heartbeat×3→completed in the mailbox. 3 zero-ref candidates = all verified false positives (string-wired framework refs: DUPEFILTER_CLASS, scrapy spider name) — no deletions; the delegates-must-verify rule proven. Runbook gains the e2m section.

coworkers-console: localhost HTML entry point to the workforce (dogfooded live)

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03. apps/coworkers-console/ per the html-effectiveness pattern (post crawled+mirrored first): single standalone index.html (inline CSS/JS, terminal-noir tokens) + stdlib serve.py (binds 127.0.0.1; /api/coworkers reads all 15 IDENTITY+BRAND, critter SVGs imported from scripts/critters.py single-source; /api/mailbox + /api/alerts = e2m envelopes with silence-alert logic; /api/reports = newest WBR/territory/sweep/toolchain). Dogfooded sequentially: built → preview_start → measured (15 cards, 5 envelopes, 4 report sections, 0 errors, critters rendering) → modal verified (directives/gates/skills/deck-line) → screenshot confirmed. Read-only v1; gated POST /api/dispatch is the recorded v2 candidate (operator decision). launch.json committed as the durable launch config. Skill in cwc-agent-resources.

Per-coworker filesystem memory with dream-only CRUD — spec-driven, all criteria pass

completed · subagentcoworkers.com · session sess_2026-07-02_wbr

DONE 2026-07-03. Operator prompt rewritten as a structured prompt per mirrored best practices (role/grounding/tasks/constraints/success-criteria/non-goals) at plugins/_schema/MEMORY-SPEC.md — which doubles as the spec. scripts/coworker_memory.py: SQLAlchemy+sqlite index & FTS5 search & lru reads over filesystem-truth stores plugins/cwc-*/memory/v<N>/; CRUD ONLY through dreams (v(N+1) new dir, inputs never mutated per dreams.md; 100KB/2000 caps per memory.md as hard errors; e2m envelopes per dream = office-agents audit telemetry; claude-tag channel-memory layer made literal). Seeded 15 coworkers via 15 real dreams. ALL FIVE success criteria verified live: v1+current ×15, v1 byte-identical after finance v2, 100001-byte put rejected with byte count, warm search 0.00ms, mailbox all-terminal incl. the failed cap-test envelope. LSP sweep re-covers the module. Follow-up recorded: sync local dreams → subagentmemories.com API when write secret provisioned; console memory rendering = v2 candidate.