colonyq@0.8.0

Minor Changes

• 86a62d9: colony bridge lifecycle gains --replay <file> and --dry-run so a saved colony-omx-lifecycle-v1 envelope (e.g. captured .pre.json) can be routed offline through the real lifecycle logic without touching the live data dir. Combined with --json, this gives runtime integrators a CI-shaped harness for asserting on route, event_type, extracted_paths, and ok.

• 99936fa: colony bridge replay <file.pre.json> is now a first-class subcommand for
  offline debugging of captured pre-tool-use envelopes. Default is --dry-run
  (ephemeral in-memory SQLite, no side effects); pass --apply to write to
  the live store. A new --rewrite-root <from>=<to> flag rewrites absolute
  paths in the envelope before dispatch so captures from another machine can
  be replayed locally. Reuses the existing
  packages/contracts/fixtures/colony-omx-lifecycle-v1/ fixtures and does not
  require the worker daemon. The shell shim at apps/cli/bin/colony.sh
  short-circuits only bridge lifecycle to the daemon, so bridge replay
  runs in-process automatically.

• edc318f: colony gain --summary now renders an rtk-style compact view over the same
  mcp_metrics receipts: headline KPI stack (total calls, input/output/total
  tokens, tokens saved, total exec time), efficiency meter, top-N By
  Operation table with proportional impact bars, a 30-day Daily Activity
  bar graph, and a 12-day Daily Breakdown table. --graph and --daily
  narrow the output to a single section; --days <n> and --top-ops <n> tune
  the window and table size. Per-operation saved-token credit is distributed
  across each comparison row's matched_operations proportionally to call share
  so the Saved column lines up with the headline total.

  Storage gains Storage.aggregateMcpMetricsDaily({ since, until, operation })
  returning per-UTC-day rollups ({ day, calls, input_tokens, output_tokens, total_tokens, total_duration_ms }) ordered newest-first. Type exports
  AggregateMcpMetricsDailyOptions and McpMetricsDailyRow come along.

• 9f6b502: Add colony grab command group: a per-project localhost intake daemon that
  turns a react-grab "Add context" submit into a colony task on a fresh
  agent/* worktree and starts a detached tmux session running codex
  inside it.

  • colony grab serve — long-lived HTTP daemon on 127.0.0.1 with strict
    request gating (bearer token, Origin allowlist, JSON content-type,
    CORS preflight). On accepted POST /grab, creates a colony task,
    posts the react-grab payload as a kind: "note" observation, writes
    .colony/INTAKE.md into the worktree, and spawns tmux new-session -d
    running codex in the worktree.
  • colony grab attach <task-id> — convenience attach to the spawned
    tmux session rg-<task-id>.
  • colony grab status — list grab daemons known to $COLONY_HOME.

  In-memory dedup (default 5 min window) keyed by
  sha256(repo_root|file_path|content|extra_prompt) collapses repeat
  submits into task_post notes on the existing task.

  The daemon is off by default; it must be started explicitly.

• 86a3d1a: colony install now auto-wires the Colony skill into the active cue profile so
  the agent discovers Colony as a pullable capability (loads on a real trigger)
  instead of relying on forced session prefaces. New colony skills wire /
  colony skills unwire subcommands shell out to cue's own cue skills add-to-profile / remove-from-profile, targeting the cue-resident
  colony/colony skill. Wiring is best-effort: a missing cue is a soft no-op that
  prints the manual npx skills add fallback (now with the recodee typo fixed),
  never a failed install. Opt out with colony install --no-skills or
  COLONY_SKILL_WIRE=0; uninstall removes the skill symmetrically.

• a83eeea: colony gain drift and a matching savings_drift_report MCP tool flag
  tools whose median tokens-per-call has drifted up or down. Default windows
  are non-overlapping: recent = last 3 days, baseline = 14 days ending 3 days
  before recent. Default thresholds: --threshold 1.25 (up), --down-threshold 0.75, --min-calls 20 per window. Classifications: up_drift,
  down_drift, new_tool (no baseline), gone (no recent), insufficient_data,
  stable.

  Storage gains Storage.mcpTokenDriftPerOperation() which computes per-operation
  medians with a ROW_NUMBER() OVER (PARTITION BY operation ORDER BY tpc)
  window function — chosen over the correlated LIMIT 1 OFFSET (COUNT-1)/2
  form because SQLite forbids outer aggregate references in scalar-subquery
  OFFSET. A mcpMetricsMinTs() helper surfaces a one-line warning when the
  baseline window starts before the first recorded metric.

• 53836ff: colony health --coach walks a repo through first-week setup. It detects
  adoption stage (fresh / installed_no_signal / early / mid_adoption)
  from cheap signals (countObservations, installed-IDE flags,
  firstObservationTs, Math.max(toolCallsSince, countMcpMetricsSince)),
  then surfaces the NEXT incomplete step from a fixed 7-step ladder:
  install_runtime → first_task_post → first_task_claim_file →
  first_task_hand_off → first_plan_claim → first_quota_release →
  first_gain_review. Each step carries an exact cmd: and tool: string.

  Progress is persisted in a new coach_progress SQLite table (migration
  014-coach-progress.ts, schema_version 13 → 14). Step completion is
  event-observed via mcp_metrics / observations, never user-clicked.
  colony gain records a coach_gain_review observation so step 7 can
  self-detect. --coach is mutually exclusive with --fix-plan and respects
  --json.

• 0950b42: ICM slice 3 — observation importance + temporal decay.

  Every observation now carries an importance tier
  (critical | high | medium | low, default medium), a rolling
  access_count, a last_accessed_at timestamp, and a weight value.
  Critical/high pin their weight to the base value and never decay;
  medium/low decay as baseWeight / (1 + access_count * 0.1) whenever
  they are read. Read paths (MemoryStore.search, getObservations,
  semanticSearch) coalesce ids into a debounced 50ms batch and flush
  the access bookkeeping in one transactional UPDATE, so heavy read
  loops trade at most one extra write per ~50ms window.

  Search and get_observations MCP responses now include importance
  and weight on each row (additive — older callers ignore them).
  task_post accepts an optional importance parameter forwarded to
  the underlying observation insert.

  New CLI subcommand colony memory prune deletes near-zero-weight
  medium/low rows; --min-weight <n> overrides the default 0.1
  threshold and --dry-run reports the candidate count without
  deleting. Critical/high are never affected.

  Storage: schema bumped to version 17 with four additive columns on
  observations and two new indexes. Storage.recordAccess,
  Storage.pruneLowDecay, and Storage.countLowDecayCandidates are
  the public primitives. (Originally targeted version 15 in isolation;
  landed at 17 alongside slice 1 memoirs and slice 2 feedback.)

• 66fa52c: Surface unpublished on-disk plan workspaces in task_plan_list, and chain plan create into plan publish

  Two related improvements so orchestrators (and fleets of codex workers) don't waste cap on a plan they have a workspace for but never registered in Colony:

  • task_plan_list now scans openspec/plans/* and merges any disk workspace whose slug is not already registered, marked registry_status: 'unpublished'. Workers cannot claim from these, but seeing them lets the orchestrator notice and run colony plan publish <slug>. Pass include_unpublished: false to mirror the legacy registered-only behavior.
  • colony plan create now accepts --publish (plus optional --publish-session, --publish-agent, --publish-auto-archive) which chains into the same publish path immediately after the workspace is created, eliminating the "I created a plan but workers don't see it" failure mode.
  • PlanInfo.registry_status gains an 'unpublished' variant.

• 61150c7: Add a Movers section to colony gain that splits the queried window into a
  trailing "recent" segment and a "prior" segment, then surfaces operations whose
  per-hour call rate, token rate, or error count has shifted materially between
  the two. Top 3 risers (▲), top 3 fallers (▼), and top 3 error risers (!) are
  listed inline above the existing Operations table. New ops (no prior activity)
  are tagged (new) and disappeared ops (gone). Two new flags: --recent-hours <n> to override the split (default: window / 7) and --no-movers to
  suppress the section. JSON output gains a live.movers payload with the same
  shape as the rendered rows.

Patch Changes

• 55581ed: Add colony demo: a 60-second guided walkthrough of file-claim contention prevention. Two simulated agents (claude-code and codex) join the same task and try to claim src/api.ts; the second agent gets blocked_active_owner, then claude-code releases and codex retries successfully. The demo runs against an isolated temp data dir and cleans up on exit, with --json for a structured transcript and --keep-data for inspection. Also ship pre-baked ~/.colony/settings.json fragments under examples/policies/ for Next.js monorepos, Python packages, and Rust workspaces — each fragment lists stack-appropriate privacy.excludePatterns (build output, caches, .env) and protected_files (lockfiles, root config). README points to both surfaces from the install block.

• 8917c73: Fix two colony health scoring bugs that surfaced as "bad" readiness areas with no real defect:

  • colony_mcp_share.mcp_tool_calls = 0 despite live MCP traffic. The counter only read tool_calls rows, missing MCP traffic when the calling agent's PostToolUse hook didn't fire for mcp__* tools. The counter now takes the max of that observed count and mcp_metrics row count (colony MCP server's own per-call receipt), with the source surfaced in source_breakdown.colony_mcp_metrics. New storage helper countMcpMetricsSince(since, until?).
  • claim_before_edit_ratio = null when any edits lacked file_path metadata. Forcing the ratio to null whenever edit_tool_calls !== edits_with_file_path turned a real 200/363 = 55% signal into a bare n/a headline. The ratio is now computed over measurable edits whenever edits_with_file_path > 0; the status field still communicates partial measurability for downstream consumers.

• e52cd83: Fix aggregateMcpMetrics error_reasons grouping so per-row counts sum to
  error_count. The grouping previously partitioned by (operation, error_code, error_message), but several handlers embed unique session IDs in their error
  messages (e.g. sub-task is claimed by codex-session-XYZ), so each race loss
  produced a distinct group. Combined with a 3-row truncation per operation, the
  result was that nearly all errors were hidden — task_plan_claim_subtask would
  report 7 errors in the Top error reasons table while the Operations table showed
  93 for the same row. Grouping now drops error_message from the key (SQLite
  picks the row with the latest ts for the sample message via its bare-column-
  with-MAX optimization) and the per-operation cap is bumped from 3 to 8 since
  codes are low-cardinality. Sum-of-reasons now matches error_count exactly.

• 9376314: Stop mislabelling generic MCP errors and reduce SQLite contention failures.

  • mcpError now codes non-TaskThreadError throws as INTERNAL_ERROR instead of OBSERVATION_NOT_ON_TASK, so validation failures and SQLite "database is locked" errors surface honestly in mcp_metrics and colony gain.
  • Storage now sets PRAGMA busy_timeout = 5000 on every connection (worker daemon, MCP server, CLI hooks all open separate handles to the same WAL DB), so concurrent writers wait the kernel out instead of throwing SQLITE_BUSY immediately.

• 5efdb52: Native Windows support for the colony CLI. The bin entry was a POSIX shell
  script (bin/colony.sh) that npm could not execute on Windows without WSL,
  breaking every Windows install of the package. The shim is now a Node ES
  module (bin/colony.mjs) using only node:* builtins, so npm's generated
  .cmd / .ps1 wrappers run it natively under cmd, PowerShell, and Git Bash.

  The daemon fast-path for colony bridge lifecycle --json is preserved — the
  HTTP POST to 127.0.0.1:$COLONY_WORKER_PORT/api/bridge/lifecycle now goes
  through node:http, with a node:net connect probe (1s) before the request
  (2s) so the fallback latency stays close to the curl-based version when the
  daemon isn't running. Stdin is buffered and replayed on fallback, preserving
  rule #10 (a dead daemon must never lose or block a write).

  CI now runs the build matrix on ubuntu-latest, macos-latest, and
  windows-latest across Node 20 and 22 so this regression cannot recur.

• 1d78c99: Push awareness: when an edit touches a file another live session holds, the PostToolUse hook now injects a one-line [Colony] session X recently claimed <file> … note into the agent's context immediately (Claude Code additionalContext), instead of waiting for the next turn's preface. Debounced to once per 2 minutes per session via an awareness-push observation marker (hook processes are one-shot, so the marker doubles as audit trail). autoClaimFromToolUse now reports live-owner blocked takeovers in its conflicts result.

• 08482b1: Surface hot-loop dominance and drop double-"saved" labels in colony gain.
  Top spend now reports the operation's share of total tokens, and a Hot loop:
  callout fires when one operation owns ≥70% of token spend across ≥100 calls.
  The "Saved:" / "USD saved:" labels are renamed to "Net:" / "Net USD:" so the
  phrase no longer reads "Saved: X saved", and the live sessions header drops the
  trailing , - when cost isn't configured.

• 658b722: Raise SQLite contention headroom in @colony/storage so the worker daemon,
  MCP server, CLI hooks, and codex-fleet panes can share ~/.colony/data.db
  without surfacing SQLITE_BUSY: database is locked to callers. The
  Storage constructor now sets busy_timeout=15000 (was 5000), and
  withBusyRetry defaults bump to 8 attempts with up-to-1s backoff (was 5
  attempts / 250ms cap). Happy-path callers are unaffected because no busy
  error still means no retry sleep; sustained contention from ~30+ concurrent
  writers — the codex-fleet shape that triggered this — now has ~3.85s of
  combined SQLite + Node retry headroom before raising.