subagenthandoffs

Expand job-board crawler with Ashby/OpenAI support, rebuild workers/etl, build subagentetl.com Kimball ETL dashboard, build cwc-data plugin, deploy all three live via Desktop Commander

ref: sessions/handoff-2026-06-30.json

None of that file's 8 tasks (task-refresh-lockfile, task-migration-006, task-migration-007, task-verify-bootstrap, task-first-crawl, task-index-docs, task-a2a-executor, task-repo-fmt-debt) were touched this session -- all Rust/Postgres/docs-crawl work, unrelated to this session's TypeScript/Wrangler/plugin scope. Presumed still pending; re-verify status rather than trusting this note indefinitely.

• HANDOFF.md — A concurrent, separate session's narrative handoff (untracked, written same day) covering disjoint work: workers/subagentworkers + workers/subagentdb (GLM-5.2 thinking-mode empty-output bug, a zero-external-paid-API product constraint, and a general 'never run npm/tsc/wrangler for any workers/* project from a cloud sandbox -- it poisons the shared node_modules mount with wrong-platform binaries' warning that applies here too, not just to those two workers). It also researched Managed Agents/Office Agents/Claude Tag independently, using this repo's local docs/docs/ mirror instead of live fetch -- conclusions converge with model_notes below, which is a useful cross-check.

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "schema_version": "0.1.0",
  "kind": "agent_handoff",
  "previous_handoff": "sessions/handoff-2026-06-30.json",
  "see_also": [
    {
      "path": "HANDOFF.md",
      "note": "A concurrent, separate session's narrative handoff (untracked, written same day) covering disjoint work: workers/subagentworkers + workers/subagentdb (GLM-5.2 thinking-mode empty-output bug, a zero-external-paid-API product constraint, and a general 'never run npm/tsc/wrangler for any workers/* project from a cloud sandbox -- it poisons the shared node_modules mount with wrong-platform binaries' warning that applies here too, not just to those two workers). It also researched Managed Agents/Office Agents/Claude Tag independently, using this repo's local docs/docs/ mirror instead of live fetch -- conclusions converge with model_notes below, which is a useful cross-check."
    }
  ],
  "meta": {
    "from_agent": "claude-sonnet-5",
    "to_agent": "claude-fable-5",
    "to_agent_note": "Fable 5 redeploys 2026-07-01 (today) per anthropic.com/news/redeploying-fable-5. Its exact API model-string slug was not published in that announcement -- confirm the real string at platform.claude.com/docs before assuming \"claude-fable-5\" is correct. If a request unexpectedly comes back from Opus 4.8 instead, that's expected: Anthropic auto-routes Fable-5 requests blocked by its new cybersecurity classifier to Opus 4.8 rather than refusing outright.",
    "repo": "opencoworkers/subagentjobs",
    "default_branch": "main",
    "created_at": "2026-07-01",
    "session_goal": "Expand job-board crawler with Ashby/OpenAI support, rebuild workers/etl, build subagentetl.com Kimball ETL dashboard, build cwc-data plugin, deploy all three live via Desktop Commander"
  },
  "mission": "Everything under completed_this_session is LIVE in production and was verified end-to-end this session (not just type-checked/assumed). Read CLAUDE.md's Pending tasks + still_open_from_previous_handoffs below before starting new work -- none of that was touched this session.",
  "completed_this_session": {
    "ashby_integration": {
      "summary": "Added Ashby ATS platform support (fetchAshbyIds) to workers/cron for OpenAI (board token 'openai'). Ingested and reconciled to 723 rows in fact_job_posting; fixed a company_name capitalization bug ('OpenAI' -> 'openai') to match the lowercase-token convention every other company uses (anthropic, figma, palantir, ...).",
      "verified": "Live crawl via deployed workers/etl: /crawl?board=openai -> {\"platform\":\"ashby\",\"changed\":true,\"jobs\":723}. Confirmed via direct D1 query."
    },
    "etl_worker_rebuild": {
      "summary": "workers/etl previously had NO in-repo source (dashboard/script-API deployed). Rebuilt src/index.ts, package.json, tsconfig.json, wrangler.toml from scratch, added fetchAshby(). Fixed a pre-existing silent dim_board bug: writes were UPDATE-only, so 43 of 45 Greenhouse boards never got a dim_board row (invisible on any job_count>0 dashboard filter despite having live job data) -- changed to INSERT ... ON CONFLICT ... DO UPDATE and backfilled the 43 boards; documented via migrations/005_expand_boards.sql + 006_backfill_dim_board_stats.sql.",
      "verified": "Post-deploy live tests: /crawl?board=openai -> 723 jobs (ashby), /crawl?board=whoop -> 164 (lever), /crawl?board=figma -> 168 (greenhouse). crawl_log rows carry real sha256 hashes, replacing the earlier manual placeholder 'manual-ingest-2026-07-01'."
    },
    "subagentetl_dashboard": {
      "summary": "New Cloudflare Worker workers/subagentetl, serves subagentetl.com. Kimball-style ETL pipeline dashboard: SVG star-schema diagram (fact_job_posting / dim_board / dim_skill / dim_department / dim_office / bridge_job_skill / bridge_job_office / crawl_log), pipeline-run history table, per-source freshness badges (STALE_HOURS=12). User explicitly chose scope: visualize the EXISTING crawler pipeline, not a new one.",
      "bugfix": "starSchemaSvg() originally hardcoded literal counts (e.g. '51 skills') as static SVG text. Fixed by threading the live D1-backed Summary object through the function so the diagram can't silently go stale.",
      "verified": "Deployed (Version 4347336a-e4e4-41a6-bb19-bfd8b0d60251). My sandboxed web_fetch showed a stale cached page post-deploy; diagnosed as a fetch-tool caching artifact (not a real regression) via cache-busted `curl -D-` run on the real machine through Desktop Commander -- response headers matched the new code exactly."
    },
    "cwc_data_plugin": {
      "summary": "New plugins/cwc-data, mirrors plugins/cwc-design's file layout exactly (.claude-plugin/plugin.json, commands/data.md, skills/data-playbook/SKILL.md, agents/data-coworker.md, gates.toml, README.md, LICENSE). Scope: warehouse analyst ONLY, per explicit user choice -- pipeline/DAG authoring and web/data-extraction skill areas deliberately excluded. Registered in .claude-plugin/marketplace.json alongside cwc-design.",
      "notes": "skills/data-playbook/SKILL.md encodes the real schema gotchas discovered this session (see gotchas array below), a 6-item self-review checklist, and a provenance-footer format, following the design basis of Anthropic's self-service-analytics blog post the user cited."
    },
    "deploys": {
      "summary": "Ran real `npm install` + `wrangler deploy` for subagentetl, workers/etl, workers/cron via the Desktop Commander MCP (operates on the user's actual Mac, not this sandbox) at the user's explicit request.",
      "bugfix": "`wrangler deploy` failed non-interactively ('More than one account available but unable to select one in non-interactive mode') because the login resolves 2 Cloudflare accounts. Fixed permanently by adding `account_id = \"e6294e3ea89f8207af387d459824aaae\"` to wrangler.toml for all three workers touched this session (cron, etl, subagentetl). Did NOT touch workers/web or other untouched workers' wrangler.toml files."
    }
  },
  "gotchas": [
    "company_name is a lowercase crawl token ('anthropic', 'openai'), never a display name.",
    "evicted_at IS NULL is the 'active' hygiene filter used everywhere.",
    "dim_department, dim_office, bridge_job_office, streams_job_change are schema-defined but empty/unused in production -- don't assume data exists there.",
    "Netflix uses Eightfold (not Greenhouse/Lever/Ashby) -- handled as a one-time manual lookup, never ingested into the warehouse.",
    "job_count = 0 on a board can be legitimate, not a bug.",
    "D1 has bound-parameter limits; the ETL's 50-row x 12-column batch upserts deliberately use manual SQL-string escaping instead of .bind() placeholders for this reason -- don't 'improve' this without checking the limit math first."
  ],
  "environment_constraints_new_this_session": [
    {
      "id": "deploy-requires-desktop-commander",
      "severity": "high",
      "fact": "This agent (cloud sandbox) has no authenticated wrangler/npm session and cannot deploy Cloudflare Workers directly.",
      "workaround": "Use the Desktop Commander MCP (mcp__Desktop_Commander__start_process / read_process_output / interact_with_process) to run npm/wrangler commands on the user's real Mac, which has an authenticated wrangler CLI. Only usable when the user explicitly invokes/grants it."
    },
    {
      "id": "wrangler-multi-account-noninteractive",
      "severity": "medium",
      "fact": "The user's Cloudflare login resolves to 2 accounts (Admin@jadecli.com and Alex@jadecli.com). `wrangler deploy` fails non-interactively with 'More than one account available' unless account_id is pinned.",
      "workaround": "Add `account_id = \"e6294e3ea89f8207af387d459824aaae\"` to any NEW worker's wrangler.toml before deploying. Already present in workers/cron, workers/etl, workers/subagentetl."
    },
    {
      "id": "connected-folder-delete-blocked",
      "severity": "low",
      "fact": "rm/mv/unlink on files inside the connected project folder fail with 'Operation not permitted' by default.",
      "workaround": "Call mcp__cowork__allow_cowork_file_delete on the path once; deletion then works for the rest of the session for that path/folder."
    }
  ],
  "still_open_from_previous_handoffs": {
    "ref": "sessions/handoff-2026-06-30.json",
    "note": "None of that file's 8 tasks (task-refresh-lockfile, task-migration-006, task-migration-007, task-verify-bootstrap, task-first-crawl, task-index-docs, task-a2a-executor, task-repo-fmt-debt) were touched this session -- all Rust/Postgres/docs-crawl work, unrelated to this session's TypeScript/Wrangler/plugin scope. Presumed still pending; re-verify status rather than trusting this note indefinitely."
  },
  "tasks": [
    {
      "id": "task-vendor-indexer-run",
      "content": "Run indexer over vendor repos (crates/indexer walks .vendors.toml repos -> fact_filesystem + dim_file_ast)",
      "priority": "medium",
      "status": "pending",
      "kind": { "type": "shell_command", "command": "cargo run -p indexer -- --all" }
    },
    {
      "id": "task-switch-cowork-profile",
      "content": "Switch Claude Desktop to the cowork profile (surface macos__desktop_cowork__engineering_coworker)",
      "priority": "low",
      "status": "pending",
      "kind": { "type": "shell_command", "command": "bash profiles/switch-profile.sh cowork" },
      "notes": "Requires building the engineering-coworker binary first: cargo build -p engineering-coworker"
    },
    {
      "id": "task-refresh-oauth-token",
      "content": "Refresh CLAUDE_CODE_OAUTH_TOKEN if expired",
      "priority": "low",
      "status": "pending",
      "kind": { "type": "shell_command", "command": "claude auth login" },
      "notes": "Conditional -- only needed if the current token has actually expired."
    }
  ],
  "model_notes": {
    "fable_5": {
      "summary": "Released 2026-06-09 alongside Mythos 5 (same underlying model; Fable 5 shipped with much stronger safeguards for general use, Mythos 5 with fewer safeguards to a small set of trusted Project Glasswing partners for defensive cybersecurity). Both hit US export controls 2026-06-12 after Amazon researchers found a technique prompting Fable 5 to identify software vulnerabilities (in one case, produce exploit code). Testing showed weaker models (incl. Opus 4.8, GPT-5.5, Kimi K2.7, and for the single-exploit case even Haiku 4.5/Sonnet 4.6/Opus 4.6/4.7) could reproduce the same outputs, so this was judged a borderline safety-margin gap, not a unique Mythos-level capability leak. Controls lifted 2026-06-30. Fable 5 redeploys 2026-07-01 (today) across Claude Platform, Claude.ai, Claude Code, and Claude Cowork. A new safety classifier blocks the specific reported technique in >99% of cases; blocked Fable-5 requests are auto-routed to Opus 4.8 instead of just refusing. Expect more false-positive blocks on routine coding/debugging while the classifier is refined. Pro/Max/Team/select-Enterprise: up to 50% of weekly usage limits through 2026-07-07, then usage credits; standard Enterprise seats need usage credits enabled or Fable 5 won't work at all.",
      "source": "https://www.anthropic.com/news/redeploying-fable-5"
    },
    "claude_managed_agents": {
      "summary": "Beta hosted agent harness (`managed-agents-2026-04-01` beta header required), a separate product from Claude Cowork (what this session runs in). Core concepts: Agent (model+prompt+tools+MCP+skills), Environment (Anthropic cloud sandbox or self-hosted), Session (running instance), Events. Explicitly stateful: sessions are long-running and 'resume cleanly after pauses', with conversation history/sandbox state/outputs persisted server-side -- the closest existing Anthropic product concept to 'resuming a session across a gap'. Not eligible for Zero Data Retention or HIPAA BAA because of that persistence. Relevant if a future session wants to move the 6h crawl cycle or ETL pipeline off Cloudflare Cron Triggers onto a hosted async agent instead.",
      "source": "https://platform.claude.com/docs/en/managed-agents/overview"
    },
    "claude_tag": {
      "summary": "'@Claude' -- a persistent, admin-governed AI teammate embedded in Slack, one shared instance per channel, runs in an ephemeral Anthropic-hosted sandbox, works with Opus 4.8. Could plausibly be used later for Slack alerts on crawl/deploy failures for this repo, but nothing wired up this session.",
      "source": "https://www.anthropic.com/news/introducing-claude-tag"
    },
    "office_agents": {
      "summary": "Resolved after initial searches turned up only adjacent products: 'Office Agents' is an admin-settings toggle within Claude for Microsoft 365 that enables cross-app context sharing -- e.g. a Word agent and an Excel agent share one conversation thread so Claude can pull live spreadsheet data while drafting a doc or building slides, without the user re-explaining context per app. Shipped 2026-03-11. Not directly relevant to this repo (no Microsoft 365 surface here); researched only because the user explicitly asked for it.",
      "source": "https://thenewstack.io/claude-word-excel-powerpoint-outlook-microsoft-office/ and https://amitkoth.com/claude-office-agents-explained/"
    }
  },
  "enum_reference": {
    "task_status": ["pending", "in_progress", "completed", "cancelled"],
    "task_priority": ["high", "medium", "low"],
    "task_kind_type": ["todo", "crawl_board", "evict_stale", "deploy", "migration", "git_push", "code_review", "sub_task", "shell_command"]
  },
  "operating_rules": [
    "Develop on a fresh branch off origin/main; never push to main.",
    "Push, then open a ready-for-review PR; check for a PR template first.",
    "End commit messages with the Co-Authored-By + Claude-Session trailers (chat-only; never reference the model id in committed artifacts).",
    "GitHub access is scoped to opencoworkers/subagentjobs only; use mcp__github__* tools (no gh CLI).",
    "Deploys in this environment require the Desktop Commander MCP against the user's real Mac -- see environment_constraints_new_this_session.deploy-requires-desktop-commander."
  ]
}

source_site subagenthandoffs.com · row created 2026-07-02 02:06:38 · JSON