bullswarm 0.9.0 → 0.10.0

package/CHANGELOG.md

@@ -1,5 +1,38 @@
 # bullswarm changelog
 
+## 0.10.0 — battle-tested advisory orchestration and agent activity
+
+- Made `maxAgents` and `maxWorkflowSeconds` advisory planning targets instead
+  of hard stops. Workflows can exceed them to finish required implementation
+  and verification; structural graph-growth limits remain enforced.
+- Removed implicit connector and generated-goal wall-clock timeouts. Delegates
+  wait for natural completion unless an operator explicitly supplies a timeout,
+  requests cancellation, or a definitive auth/quota failure is observed.
+- Added compatibility migration for generated 0.9.0 goals carrying Bullswarm's
+  former 900-second planner/action timeout defaults.
+- Fixed adaptive completion policy, current-action metadata, provider routing
+  history, usage aggregation, latest-worker verification, and truthful partial
+  token/cost accounting found during the Kipwise battle test.
+- Added connector-owned native JSONL event adapters for Codex, Claude, Grok,
+  Command Code, and OpenCode. Workflows now retain and display the latest three
+  semantic shell/read/edit/write/response actions for every active agent.
+- Added conservative stall evidence: ten minutes without transport, parsed
+  event, or semantic action activity is labeled `suspected_stalled` but never
+  causes an automatic kill.
+- Added a low-noise `workflow watch <id>` progress stream with semantic-change
+  updates, heartbeats, last-three agent actions, and terminal per-attempt timing.
+- Added optional durable `workflow steer <id> --message ...` guidance delivered
+  only at the next planning checkpoint, never injected into an active worker.
+- Captured runtime model IDs declared in provider event streams for more complete
+  model/cost attribution, preserved Grok tool kinds across name-less updates,
+  and enabled Claude's supported forwarded-subagent text stream.
+- Taught planners to reuse clean full-suite evidence and isolate mutation/pre-fix
+  experiments instead of redundantly or concurrently testing a changing tree.
+- Expanded the offline suite to 224 tests. Real bounded CLI probes confirmed
+  all five provider event formats, and an exact packaged OpenCode watch smoke
+  passed argument injection, action normalization, final-output extraction,
+  and the content gate together.
+
 ## 0.9.0 — resilient dynamic workflow routing
 
 - Added cooperative `SIGTERM`/`SIGINT` handling, durable `interrupted` states,

package/README.md

@@ -114,6 +114,7 @@ The detached response includes a short ID and exact observation commands:
 
 ```bash
 bullswarm workflow runs show <shortId>
+bullswarm workflow watch <shortId>        # low-noise live progress + terminal timing
 bullswarm workflow tui <shortId>          # printable phase/action/attempt tree
 bullswarm workflow tui --json <shortId>
 bullswarm workflow events --json <shortId> --after 0
@@ -183,6 +184,18 @@ bullswarm workflow run audit-code --resume <shortId>
 
 ### Live workflow dashboard
 
+For ordinary observation, use the non-interactive watcher. It prints only when
+the phase, step, agent action, routing, or status changes, plus a 60-second
+heartbeat while otherwise quiet. Terminal output includes every attempt's
+agent/model, elapsed time, outcome, and tokens, so a slow test is distinguishable
+from a stalled process without writing a polling script.
+
+```bash
+bullswarm workflow watch <shortId>
+bullswarm workflow watch <shortId> --jsonl       # automation-friendly stream
+bullswarm workflow watch <shortId> --once        # one current/terminal snapshot
+```
+
 `workflow tui` is the interactive, Claude-style `/workflows` view. It watches
 ongoing runs from disk and supports `j`/`k` or arrow-key selection, Enter for
 details, `c` to request a cooperative stop, `r` to refresh, and `q` to quit.
@@ -202,6 +215,7 @@ bullswarm workflow tui --json --cancel <id>   # request cooperative stop
 bullswarm workflow capabilities --json       # pools, lanes, models, meters, limits
 bullswarm workflow inspect <file-or-name>     # workflow shape and semantics
 bullswarm workflow events --json <id> --after 20
+bullswarm workflow steer <id> --message "Prefer focused tests before another full suite"
 bullswarm workflow action show --json <id> <actionId>
 bullswarm workflow approval approve --json <id>  # then resume the run
 ```
@@ -213,16 +227,50 @@ commit a distinct resumable `interrupted` state. On every workflow command,
 active states with a dead/stale owner are automatically reconciled to
 `interrupted` instead of remaining falsely `running`.
 
+`workflow steer` is optional operator guidance, not hot-patching. It appends a
+durable instruction that is delivered only to the next not-yet-started
+`decide` checkpoint; the active worker continues unchanged. Steering remains
+inside the original goal and authorization boundary and cannot bypass runtime
+validation or required verification. Static workflows and terminal runs reject
+steering because they have no future orchestration checkpoint.
+Live attempts record the last stdout/stderr activity time and observed byte
+count separately from the runner heartbeat. This makes a silent process
+visible without treating elapsed wall time alone as proof that it is hung.
+Supported coding-agent connectors also enable their native JSONL event mode and
+declaratively map provider events into a common semantic action record:
+
+```json
+{"id":"provider-action-id","at":"...","kind":"shell_command|read_file|edit|response","status":"running|completed|failed","summary":"safe scalar preview"}
+```
+
+The live workflow pane retains the latest three logical actions per agent.
+Repeated updates for the same tool call replace its status, and streaming text
+chunks coalesce into one response action. Heartbeats, token/thought deltas,
+usage messages, hooks, and unparsed output remain liveness evidence but do not
+occupy the action pane. Connector-specific flags, paths, and mappings live in
+`connectors/*.json` under `eventStream`; core contains no provider event names.
+
+After ten minutes without transport, parsed-event, or semantic-action evidence,
+an active child is labeled `suspected_stalled`. This is an inspection signal,
+not a death verdict and never an automatic kill: buffered CLIs can be silent
+while working. Process exit, a fatal auth/quota signature, explicit operator
+cancellation, or an opt-in timeout remain the terminal signals.
+
 Each attempt records the phase/action, selected pool and model, effort tier,
 routing reason, all eligible candidates with quota surplus, timestamps,
 artifact paths, outcome, and reported-or-estimated token/cost/quota usage.
 `workflow tui <id>` renders this breakdown for completed runs as well as live
 ones; `workflow tui --json <id>` exposes the durable audit document.
+When a provider event stream reports the actual model, Bullswarm records that
+runtime value and uses its matching connector rate metadata for the attempt's
+cost estimate. Unknown or provider-hidden model identity remains explicitly
+unknown.
 
 ### Adaptive workflows
 
 Static workflows remain zero-extra-LLM orchestration. An adaptive workflow adds
-an explicit `decide` step and hard limits:
+an explicit `decide` step, advisory resource targets, and structural expansion
+limits:
 
 ```json
 {
@@ -249,6 +297,13 @@ an explicit `decide` step and hard limits:
 }
 ```
 
+`maxAgents` and `maxWorkflowSeconds` are advisory inputs to the orchestrator.
+Crossing either target is recorded in durable state but never stops a worker,
+skips verification, or fails a run. `maxExpansionRounds`, `maxActions`, and
+`maxItemsPerExpansion` remain hard graph-growth safeguards. Delegates have no
+implicit wall-clock timeout; set a step's `timeoutSec` (or direct-run
+`--timeout`) only when an operator explicitly wants a hard termination timer.
+
 The planner returns versioned JSON. It may propose `needs_more_work` with
 bounded `run`, inline-`fanout`, or `verify` actions. The deterministic runtime
 validates IDs, dependencies, operation types, capabilities, and budgets before

package/connectors/_schema.json

@@ -12,9 +12,35 @@
   "authSignatures": ["strings in output that mean auth/throttle failure"],
   "outputExtraction": {
     "$comment": "how to get the real answer out of stdout+stderr+files",
-    "strategy": "stdout|stdout-tail|json-field|file",
+    "strategy": "stdout|stdout-tail|json-field|file|event-stream",
     "field": "optional json field path or file glob"
   },
+  "eventStream": {
+    "$comment": "Optional declarative JSONL adapter. Provider quirks stay here; core emits normalized semantic actions.",
+    "format": "jsonl",
+    "args": ["CLI flags that enable machine-readable streaming"],
+    "modelPaths": ["provider model-id paths in priority order"],
+    "silenceThresholdSec": 600,
+    "rules": [{
+      "rootMatch": {"path": "provider.event.type", "equals": "tool_started"},
+      "forEach": "optional.array.path",
+      "match": {"path": "nested.type", "equals": "tool_use"},
+      "idPaths": ["provider action id paths in priority order"],
+      "kindPaths": ["tool or response kind paths"],
+      "summaryPaths": ["safe command/path/text scalar paths"],
+      "status": "queued|running|streaming|completed|failed",
+      "statusPath": "optional provider status path",
+      "statusMap": {"provider-status": "normalized-status"},
+      "aggregate": "consecutive",
+      "summaryMode": "replace|concat"
+    }],
+    "output": [{
+      "match": {"path": "provider.event.type", "equals": "final"},
+      "path": "provider final response scalar path",
+      "mode": "last|concat",
+      "separator": "optional separator"
+    }]
+  },
   "meter": {
     "$comment": "none | declared (manual) | reader (programmatic)",
     "type": "none|declared|reader",
@@ -50,5 +76,5 @@
     "stealth": false,
     "$comment-stealth": "stealth=true: prompts/completions retained by an anonymous provider — opt-in only, never a default"
   },
-  "timeoutSec": 900
+  "$comment-timeout": "Delegates have no implicit wall-clock timeout; use direct --timeout or workflow step.timeoutSec as an explicit opt-in"
 }

package/connectors/claude-code.json

@@ -23,7 +23,21 @@
     "please run /login"
   ],
   "outputExtraction": {
-    "strategy": "stdout"
+    "strategy": "event-stream"
+  },
+  "eventStream": {
+    "format": "jsonl",
+    "args": ["--output-format", "stream-json", "--verbose", "--forward-subagent-text"],
+    "modelPaths": ["model", "message.model"],
+    "silenceThresholdSec": 600,
+    "rules": [
+      { "rootMatch": { "path": "type", "equals": "assistant" }, "forEach": "message.content", "match": { "path": "type", "equals": "tool_use" }, "idPaths": ["id"], "kindPaths": ["name"], "summaryPaths": ["input.command", "input.file_path", "input.path", "input.pattern"], "status": "running" },
+      { "rootMatch": { "path": "type", "equals": "user" }, "forEach": "message.content", "match": { "path": "type", "equals": "tool_result" }, "idPaths": ["tool_use_id"], "kind": "tool", "statusPath": "is_error", "statusMap": { "true": "failed", "false": "completed" }, "defaultStatus": "completed" },
+      { "rootMatch": { "path": "type", "equals": "assistant" }, "forEach": "message.content", "match": { "path": "type", "equals": "text" }, "kind": "response", "summaryPaths": ["text"], "status": "completed" }
+    ],
+    "output": [
+      { "match": { "path": "type", "equals": "result" }, "path": "result", "mode": "last" }
+    ]
   },
   "$comment-meter": "the CALLER pool \u2014 competes in analyze/build, wins only when no delegate can take the lane",
   "meter": {
@@ -51,5 +65,5 @@
     "stealth": false,
     "isCaller": true
   },
-  "timeoutSec": 900
+  "$comment-timeout": "No implicit wall-clock timeout; callers may opt in with --timeout or workflow step.timeoutSec"
 }

package/connectors/codex.json

@@ -22,7 +22,19 @@
     "invalid api key"
   ],
   "outputExtraction": {
-    "strategy": "stdout"
+    "strategy": "event-stream"
+  },
+  "eventStream": {
+    "format": "jsonl",
+    "args": ["--json"],
+    "silenceThresholdSec": 600,
+    "rules": [
+      { "rootMatch": { "path": "type", "equals": "item.started" }, "idPaths": ["item.id"], "kindPaths": ["item.type"], "kindMap": { "agent_message": "response" }, "summaryPaths": ["item.command", "item.text"], "status": "running" },
+      { "rootMatch": { "path": "type", "equals": "item.completed" }, "idPaths": ["item.id"], "kindPaths": ["item.type"], "kindMap": { "agent_message": "response" }, "summaryPaths": ["item.command", "item.text"], "status": "completed" }
+    ],
+    "output": [
+      { "match": { "path": "type", "equals": "item.completed" }, "path": "item.text", "mode": "last" }
+    ]
   },
   "meter": {
     "type": "reader",
@@ -49,5 +61,5 @@
   "flags": {
     "stealth": false
   },
-  "timeoutSec": 900
+  "$comment-timeout": "No implicit wall-clock timeout; callers may opt in with --timeout or workflow step.timeoutSec"
 }

package/connectors/command-code.json

@@ -22,7 +22,23 @@
     "cmd login"
   ],
   "outputExtraction": {
-    "strategy": "stdout"
+    "strategy": "event-stream"
+  },
+  "eventStream": {
+    "format": "jsonl",
+    "args": ["--output-format", "json"],
+    "modelPaths": ["event.model", "model"],
+    "silenceThresholdSec": 600,
+    "rules": [
+      { "rootMatch": { "path": "event.type", "equals": "tool_queued" }, "idPaths": ["event.toolCallId"], "kindPaths": ["event.toolName"], "summaryPaths": ["event.input.command", "event.input.file_path", "event.input.path"], "status": "queued" },
+      { "rootMatch": { "path": "event.type", "equals": "tool_running" }, "idPaths": ["event.toolCallId"], "kindPaths": ["event.toolName"], "summaryPaths": ["event.description"], "status": "running" },
+      { "rootMatch": { "path": "event.type", "equals": "tool_completed" }, "idPaths": ["event.toolCallId"], "kindPaths": ["event.toolName"], "status": "completed" },
+      { "rootMatch": { "path": "event.type", "equals": "message_end" }, "forEach": "event.content", "match": { "path": "type", "equals": "text" }, "kind": "response", "summaryPaths": ["text"], "status": "completed" }
+    ],
+    "output": [
+      { "match": { "path": "type", "equals": "result" }, "path": "finalText", "mode": "last" },
+      { "match": { "path": "event.type", "equals": "run_end" }, "path": "event.result.finalText", "mode": "last" }
+    ]
   },
   "meter": {
     "type": "reader",
@@ -53,5 +69,5 @@
   "flags": {
     "stealth": false
   },
-  "timeoutSec": 900
+  "$comment-timeout": "No implicit wall-clock timeout; callers may opt in with --timeout or workflow step.timeoutSec"
 }

package/connectors/echo-worker.mjs

@@ -8,7 +8,14 @@
 import { readFileSync } from 'node:fs';
 
 const task = readFileSync(process.argv[2], 'utf8');
+const sleepMatch = task.match(/SLEEP_MS:(\d+)/);
+if (sleepMatch) await new Promise((resolve) => setTimeout(resolve, Number(sleepMatch[1])));
 
+if (task.includes('FAIL:auth-hang')) {
+  console.log('Authentication failed: quota exhausted; waiting process should be terminated.');
+  await new Promise((resolve) => setTimeout(resolve, 5000));
+  process.exit(0);
+}
 if (task.includes('FAIL:auth')) {
   console.log('Authentication failed: no credentials found in keychain.');
   process.exit(0);

package/connectors/echo.json

@@ -24,5 +24,5 @@
     "pricingUpdatedAt": "2026-08-27"
   }],
   "flags": { "stealth": false },
-  "timeoutSec": 120
+  "$comment-timeout": "No implicit wall-clock timeout; callers may opt in with --timeout or workflow step.timeoutSec"
 }

package/connectors/grok.json

@@ -18,7 +18,20 @@
     "invalid api key"
   ],
   "outputExtraction": {
-    "strategy": "stdout"
+    "strategy": "event-stream"
+  },
+  "eventStream": {
+    "format": "jsonl",
+    "args": ["--output-format", "streaming-json"],
+    "silenceThresholdSec": 600,
+    "rules": [
+      { "rootMatch": { "path": "type", "equals": "tool_call" }, "idPaths": ["toolCallId"], "kindPaths": ["toolName", "title"], "summaryPaths": ["rawInput.command", "rawInput.file_path", "rawInput.path", "title"], "statusPath": "status" },
+      { "rootMatch": { "path": "type", "equals": "tool_call_update" }, "idPaths": ["toolCallId"], "kindPaths": ["toolName"], "summaryPaths": ["rawOutput.command"], "statusPath": "status" },
+      { "rootMatch": { "path": "type", "equals": "text" }, "kind": "response", "summaryPaths": ["data"], "status": "streaming", "aggregate": "consecutive", "summaryMode": "concat" }
+    ],
+    "output": [
+      { "match": { "path": "type", "equals": "text" }, "path": "data", "mode": "concat" }
+    ]
   },
   "meter": {
     "type": "reader",
@@ -43,6 +56,6 @@
   "flags": {
     "stealth": false
   },
-  "timeoutSec": 900,
+  "$comment-timeout": "No implicit wall-clock timeout; callers may opt in with --timeout or workflow step.timeoutSec",
   "$comment-meter": "weekly shared credit pool via billing endpoint; no 5h window exists on unified-billing accounts"
 }

package/connectors/opencode2.json

@@ -8,7 +8,19 @@
     "$comment-cwdMode": "QUIRK: resolves its project from $PWD, not the spawn cwd. The watcher MUST set env.PWD and spawn with cwd inside the target repo, or it will silently analyse the wrong repository and answer confidently about it."
   },
   "authSignatures": ["No cookie auth credentials found", "unauthorized"],
-  "outputExtraction": { "strategy": "stdout" },
+  "outputExtraction": { "strategy": "event-stream" },
+  "eventStream": {
+    "format": "jsonl",
+    "args": ["--format", "json"],
+    "silenceThresholdSec": 600,
+    "rules": [
+      { "rootMatch": { "path": "type", "equals": "tool_use" }, "idPaths": ["part.callID", "part.id"], "kindPaths": ["part.tool"], "summaryPaths": ["part.state.input.command", "part.state.input.file_path", "part.state.input.path", "part.state.title"], "statusPath": "part.state.status" },
+      { "rootMatch": { "path": "type", "equals": "text" }, "idPaths": ["part.id"], "kind": "response", "summaryPaths": ["part.text"], "status": "completed" }
+    ],
+    "output": [
+      { "match": { "path": "type", "equals": "text" }, "path": "part.text", "mode": "concat", "separator": "\n" }
+    ]
+  },
   "$comment-auto": "--auto is required for headless workflow dispatch: task files live under ~/.bullswarm, outside the target repo, and OpenCode otherwise pauses for an interactive permission approval. --model pins the QA/runtime pool to Luna instead of the CLI default.",
   "$comment-exit1": "known failure mode: writes a complete correct answer, then dies with a Console-sync auth error and exit 1. The verdict sets contentUsableDespiteExit instead of discarding the work.",
   "meter": { "type": "none" },
@@ -24,5 +36,5 @@
   ],
   "subscription": { "plan": null, "monthlyPriceUsd": null, "includedValueUsd": null, "quotaWindow": null },
   "flags": { "stealth": false },
-  "timeoutSec": 900
+  "$comment-timeout": "No implicit wall-clock timeout; callers may opt in with --timeout or workflow step.timeoutSec"
 }

package/docs/experiments/2026-08-28-trending-ai-autonomy.md

@@ -0,0 +1,208 @@
+# Trending AI repository autonomy experiment — 2026-08-28
+
+## Question
+
+Can the current Bullswarm source checkout autonomously plan, delegate, implement,
+and verify real work in unfamiliar, currently trending AI repositories without
+operator intervention after launch?
+
+## Frozen protocol
+
+- Research snapshot: GitHub Trending and live repository metadata on 2026-08-28.
+- Candidate shortlist: `tt-a1i/archify`, `HKUDS/nanobot`,
+  `thedotmack/claude-mem`, and `bilawalsidhu/gods-eye-view`.
+- Execution subjects: Archify and nanobot. They have local, credential-free test
+  surfaces and represent JavaScript and Python agent-oriented projects.
+- Each subject is a clean, shallow clone at the commit recorded below.
+- Dependency installation and baseline tests occur before workflow launch.
+- After launch, the operator may only read Bullswarm state/events/logs and target
+  repository state. No prompt correction, follow-up action, edit, reroute,
+  cancellation, retry, or manual repair is allowed before terminal status.
+- After terminal status, independent read-only verification may run tests and
+  inspect diffs. It may not repair the result.
+- A run counts as autonomous success only when its durable workflow reaches
+  `completed`, the requested repository change exists, focused tests pass, the
+  original acceptance criteria are met by content, and intervention count is 0.
+
+## Frozen subjects and goals
+
+### Archify
+
+- Repository: <https://github.com/tt-a1i/archify>
+- Baseline commit: `49a7821d194a70c531219f48fd0d6a08ba9ba9d7`
+- Baseline: `npm test` from `archify/` passed 721 tests with 25 skips (746 total).
+- Goal:
+
+> Work autonomously in this repository. Inspect the core non-network Archify CLI,
+> validators, renderers, and tests; identify one concrete correctness bug that is
+> not already covered by an existing test; reproduce it locally; implement the
+> smallest safe fix; add a focused regression test; run the focused test and the
+> full relevant suite; then have an independent skeptical verifier review the diff
+> and evidence. Do not use external credentials or services, do not change generated
+> release artifacts unless repository checks require it, and do not push or open a
+> PR. Do not invent a bug or make a cosmetic/docs-only change: if no defensible bug
+> can be proven, finish honestly with evidence instead of editing.
+
+### nanobot
+
+- Repository: <https://github.com/HKUDS/nanobot>
+- Baseline commit: `29025f5a8bfaeed8a8c0daf22c770afd9d023dd0`
+- Baseline: 6,257 passed, 24 skipped, 1 pre-existing unrelated failure in
+  `tests/cli/test_tui_launcher.py::test_launcher_keeps_the_tui_alive_while_an_existing_gateway_recovers`.
+- Issue: <https://github.com/HKUDS/nanobot/issues/5428>
+- Goal:
+
+> Autonomously implement HKUDS/nanobot issue #5428 in this checkout: AgentLoop
+> retains empty active-task groups after session tasks finish. First inspect the
+> repository instructions and reproduce the issue. Make the smallest architecture-
+> compliant fix and focused regression tests proving: the key disappears after the
+> only task completes; it remains until all tasks in the same group complete; and an
+> old group callback cannot delete a replacement group for the same session. Run
+> focused tests, ruff on touched Python, and the relevant/full suite as practical,
+> preserving and distinguishing the known pre-existing TUI launcher baseline failure.
+> Require an independent skeptical verifier to review the diff and evidence. Do not
+> use external credentials/services, push, or open a PR.
+
+## Results
+
+### Outcome
+
+Both frozen workflows completed autonomously. Operator intervention after each
+launch was **0**: no prompt corrections, follow-up messages, edits, retries,
+reroutes, cancellations, or repairs were made. The observer only read workflow
+events/state and, after terminal status, ran independent read-only verification.
+
+| Subject | Run | Result | Elapsed | Dispatches | Adaptive rounds | Known tokens | Cost visibility |
+| --- | --- | --- | ---: | ---: | ---: | ---: | --- |
+| Archify | `wf-mtchnk88-e8691f` (`awd62a`) | completed | 28m 43s | 5 | 1 | 19,021 | $0.03708 known subtotal; Claude portion unknown |
+| nanobot | `wf-mtcipy51-ca04c8` (`becfki`) | completed | 46m 05s | 7 | 2 | 31,910 | unknown; Claude and Command Code model rates unavailable |
+
+Both runs stayed below the advisory 30-dispatch and 60-minute planning targets.
+Those values were exposed to the orchestrator as planning context and did not
+hard-stop or skip any action.
+
+### Archify execution
+
+Bullswarm initially selected Grok 4.6 because its live weekly meter had the most
+surplus. The orchestrator created one cohesive discovery/implementation action,
+followed by a dependent skeptical verification action and a final decision gate.
+
+| Phase / step | Agent | Status | Known tokens | Evidence |
+| --- | --- | --- | ---: | --- |
+| Plan 1 / `orchestrator` | Grok 4.6 | succeeded | 3,415 | created bounded find/fix plus verify plan |
+| Execute / `find-fix-regression` | Grok 4.6 | succeeded | 1,874 | found, reproduced, fixed, and tested a real delta-reporting bug |
+| Verify / `verify-find-fix` | Grok 4.6 | succeeded | 1,597 | independently accepted the content and tests |
+| Final gate / `orchestrator` attempt 2 | Grok 4.6 | retryable failure | 5,794 | provider emitted a rate-limit signature |
+| Final gate / `orchestrator` attempt 3 | Claude Code | succeeded | 6,341 | Bullswarm quarantined Grok and rerouted without operator help |
+
+The discovered bug was that the architecture delta's canonical hash included
+`components[].brand`, while the semantic component-field list omitted `brand`.
+A brand-only edit therefore changed the canonical hash but reported zero changed
+components and produced no navigator row. The autonomous result:
+
+- adds `brand` to `COMPONENT_FIELDS.semantic` in
+  `archify/delta/architecture-delta.mjs`;
+- adds a focused 22-line regression test;
+- rebuilds `archify.zip` because the repository's package-freshness gate requires
+  it; and
+- leaves a worker evidence report, with no commit, push, or PR.
+
+The new test failed against the original implementation (18/19 passing), then
+passed with the fix (19/19). The full suite finished at 747 total, 722 passed,
+25 skipped, 0 failed. Independent post-terminal verification repeated the 19
+focused tests and `git diff --check`; both passed.
+
+### nanobot execution
+
+Grok remained quarantined after the Archify rate limit, so Bullswarm selected
+Claude Code for orchestration and Command Code for implementation/verification.
+
+| Phase / step | Agent | Elapsed | Status | Known tokens | Evidence |
+| --- | --- | ---: | --- | ---: | --- |
+| Plan 1 / `orchestrator` | Claude Code | 4m 26s | succeeded | 4,970 | localized issue #5428 and planned fix plus skeptical verify |
+| Execute / `implement-fix` | Command Code | 9m 16s | succeeded | 2,326 | implemented pruning callback and initial tests |
+| Verify 1 / `verify-fix` | Command Code | 6m 34s | succeeded | 2,249 | returned a pass, but its evidence was later challenged |
+| Gate 2 / `orchestrator` | Claude Code | 5m 28s | succeeded | 8,652 | mutation-tested the result, rejected the weak pass, expanded plan |
+| Execute / `strengthen-tests` | Command Code | 12m 39s | succeeded | 2,036 | repaired identity test and added real `AgentLoop.run()` coverage |
+| Verify 2 / `verify-strengthened-tests` | Command Code | 5m 08s | succeeded | 1,716 | independently killed both mutations and ran focused/full suites |
+| Final gate / `orchestrator` | Claude Code | 2m 34s | succeeded | 9,961 | spot-checked durable state and declared complete |
+
+The implementation adds `_prune_active_task(key, group, task)` and binds the
+specific key and group with `functools.partial`. It removes a completed task,
+drops the dictionary key only after the bound group becomes empty, and uses an
+identity guard so a late callback cannot remove a replacement group. The final
+diff is two files, 235 insertions and 1 deletion: 21/-1 production lines and 214
+test lines.
+
+The important autonomous behavior was the second planning round. Although the
+first verifier returned a pass, the Claude gate mutated the code and found that:
+
+1. removing the group-identity guard still passed the three initial tests; and
+2. restoring the original buggy callback wiring still passed the relevant tests.
+
+Bullswarm therefore did not complete. It dispatched a test-strengthening worker
+and a second verifier. The strengthened tests then killed both mutations: removing
+the identity guard fails the replacement-group test, and restoring the original
+callback fails the real `AgentLoop.run()` dispatch-path test.
+
+Independent post-terminal checks confirmed:
+
+- 4/4 strengthened pruning tests pass;
+- Ruff passes on both touched Python files;
+- `git diff --check` passes;
+- the stash is empty and only the intended two files are modified; and
+- no commit, push, PR, credentials, or external services were used.
+
+The full suite was run post-fix by autonomous workers and reported 5,337 passed,
+8 skipped, and the same single pre-existing TUI launcher failure. The clean
+baseline independently demonstrated that failure before launch; it is unrelated
+to the active-task change.
+
+### Provider-meter observations
+
+Live Fleetlens snapshots before and after the two runs showed Codex weekly usage
+moving from 28% to 30%, Claude from 49% to 50%, and Grok from 26% to 31%.
+Command Code remained at 42.0522% weekly / 40.57 credits remaining; its meter did
+not expose a visible delta at this granularity. These snapshots are useful quota
+signals, not precise per-workflow billing attribution.
+
+### Friction and remaining gaps
+
+The autonomous delivery behavior passed, but the audit exposed four reporting or
+efficiency gaps:
+
+1. Claude Code and Command Code attempts record `model: null`; consequently their
+   dollar costs cannot be calculated and both run totals are partial.
+2. Claude's internal "advisor" activity is visible only as prose in its stream,
+   not as a separately attributable Bullswarm agent or usage row.
+3. Some Grok tool events normalize to `kind: null`, even though byte activity and
+   neighboring read/search actions remain visible.
+4. One verifier's early full-suite evidence was tainted by temporarily stashing
+   tracked files while its test process was active. That evidence was excluded;
+   clean worker, later verifier, baseline, and independent checks supplied the
+   accepted proof. The workflow could more explicitly isolate mutation/pre-fix
+   checks from concurrent long-running tests.
+
+Post-experiment follow-up adds `workflow watch` for low-noise progress and
+terminal timing, captures actual Claude/Command Code model IDs when their event
+streams expose them, preserves Grok tool kinds across name-less update frames,
+queues optional steering only at future planning checkpoints, and tells the
+planner to reuse clean full-suite evidence and isolate mutation/pre-fix checks.
+Claude's forwarded-subagent text is now requested for better activity context,
+but provider-internal advisor usage still cannot be separately attributed unless
+the provider exposes distinct model/usage records in the outer event stream.
+
+Long quiet full-suite periods did not produce false dead-agent decisions: the
+event stream showed the last shell action, increasing elapsed silence, and an
+active process/output-byte signal until the command finished.
+
+## Conclusion
+
+For these two real, unfamiliar, locally testable AI repositories, Bullswarm met
+the frozen autonomy bar: it planned dynamically, delegated work, verified by
+content, rerouted around a provider rate limit, rejected a false-positive test
+verdict, expanded its plan, and reached completion with zero operator steering.
+This is strong evidence for autonomous bounded repository work, not a claim that
+every repository or credentialed/browser/deployment workflow is solved. Model and
+internal-subagent attribution remain the clearest gaps before calling the audit
+and spend breakdown perfect.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "bullswarm",
-  "version": "0.9.0",
+  "version": "0.10.0",
   "description": "Route work across coding-agent CLI subscriptions — paced by live quota meters, verified by content, never trusting exit codes.",
   "type": "module",
   "bin": {