@getpipher/armory-fleet 0.12.0 → 0.12.2

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -1,31 +1,388 @@
1
+ <div align="center">
2
+
3
+ <img src="assets/armory-fleet-logo.png" alt="@getpipher/armory-fleet" width="220" height="220" />
4
+
1
5
  # @getpipher/armory-fleet
2
6
 
3
- > The armory suite's subagent orchestrator for the [pi coding agent](https://github.com/earendil-works/pi-coding-agent) — a cross-harness, superpowers-native fleet where every agent is **armory-native from birth**.
7
+ **The armory suite's subagent orchestrator for the [pi coding agent](https://github.com/earendil-works/pi-coding-agent) — a cross-harness, superpowers-native fleet where every agent is armory-native from birth.**
8
+
9
+ [![npm version](https://img.shields.io/npm/v/@getpipher/armory-fleet?color=cb3837&logo=npm&label=npm)](https://www.npmjs.com/package/@getpipher/armory-fleet)
10
+ [![npm downloads](https://img.shields.io/npm/dm/@getpipher/armory-fleet?color=cb3837&logo=npm)](https://www.npmjs.com/package/@getpipher/armory-fleet)
11
+ [![pi compatibility](https://img.shields.io/badge/pi-%5E0.81.1-6f42c1?logo=data:image/svg+xml;base64,PHN2ZyB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciIHZpZXdCb3g9IjAgMCAyNCAyNCI+PHRleHQgeD0iNCIgeT0iMTgiIGZvbnQtc2l6ZT0iMTYiIGZpbGw9IiNmZmYiPuKCrTwvdGV4dD48L3N2Zz4=)](https://github.com/earendil-works/pi-coding-agent)
12
+ [![license](https://img.shields.io/github/license/getpipher/armory-fleet?color=blue)](./LICENSE)
13
+ [![tests](https://img.shields.io/badge/tests-593%20passing-21c463?logo=jest)](#testing)
14
+ [![release](https://img.shields.io/github/v/release/getpipher/armory-fleet?color=success&label=latest%20release)](https://github.com/getpipher/armory-fleet/releases)
15
+ [![platform](https://img.shields.io/badge/platform-macOS%20%7C%20Linux%20%7C%20WSL-lightgrey)](#compatibility)
4
16
 
5
- `@getpipher/armory-fleet` brings Claude-Code-style autonomous sub-agents to pi, built so that the armory ecosystem ([armory-todo](https://github.com/getpipher/armory-todo), [armory-memory](https://github.com/getpipher/armory-memory), [vision](https://github.com/getpipher/vision), [cursor](https://github.com/getpipher/cursor)) is the *default substrate* agents run on — not a bolt-on.
17
+ [Why](#why) · [Features](#features-at-a-glance) · [Quick start](#quick-start) · [Architecture](#architecture) · [The fleet panel](#the-fleet-panel) · [Workflows-as-code](#workflows-as-code) · [Cost-aware tiers](#cost-aware-tiers) · [Roadmap](#roadmap) · [Ecosystem](#ecosystem)
18
+
19
+ </div>
20
+
21
+ ---
6
22
 
7
23
  ## Why
8
24
 
9
- The pi-subagent ecosystem is crowded (nicobailon/pi-subagents at 2688⭐, tintinweb at 702⭐, QuintinShaw at 287⭐, and more). Three structural gaps remain open:
25
+ The pi-subagent ecosystem is crowded — `nicobailon/pi-subagents` (2688⭐), `tintinweb/pi-subagents` (702⭐), `QuintinShaw/pi-dynamic-workflows` (287⭐), `kky42/pi-flow` (66⭐), `teelicht/pi-superagents` (54⭐). Three structural gaps remain open, and `armory-fleet` owns all three — then reaches parity on the rest to be the best pi-subagent package in the ecosystem.
26
+
27
+ | Gap | Status quo | armory-fleet |
28
+ |---|---|---|
29
+ | **Armory-native integration** | No external package can integrate with the armory suite; they don't own it. | Agents that sync to [armory-todo](https://github.com/getpipher/armory-todo), hydrate from [armory-memory](https://github.com/getpipher/armory-memory), see via [vision](https://github.com/getpipher/vision), and edit via [cursor](https://github.com/getpipher/cursor) **by default** — uncopyable. |
30
+ | **Cross-harness peers** | Only one early attempt runs Claude Code + Pi as peer backends, foreground-only. | First-class dual-arsenal topology — pi and Claude Code spawn as sibling backends from one fleet. |
31
+ | **Superpowers-native lifecycle** | Only one attempt wraps the superpowers skill pipeline, synchronous-only. | The full superpowers lifecycle (brainstorm → plan → implement → review → finish) with checkpoints, quality gates, and lifecycle hooks baked in. |
32
+
33
+ Beyond those: a fleet TUI, cron/interval scheduling, git worktree isolation, cost accounting, quality gates, workflows-as-code, and a journaled event-bus — all in one package.
34
+
35
+ > **Vision, spine, and the 7-SPEC roadmap** live in [`PRD.md`](./PRD.md). The landscape deep-read (11+ packages mapped, 5 contenders deep-read) is in [`research/`](./research).
36
+
37
+ ---
38
+
39
+ ## Features at a glance
40
+
41
+ | Capability | What you get |
42
+ |---|---|
43
+ | 🧬 **Armory-native agents** | Every child syncs to `armory-todo`, hydrates from `armory-memory`, sees via `vision`, edits via `cursor` — by default, from birth. No bolt-on. |
44
+ | 🏛️ **Cross-harness backends** | Spawn pi **and** Claude Code sessions as peer backends from one fleet. Auto-detect Claude; hook-parity keeps both on equal footing. |
45
+ | 🦸 **Superpowers lifecycle** | `brainstorm → plan → implement → review → finish`, checkpoint-driven, skill-loaded per phase. `/fleet-implement <task>` runs the whole pipeline. |
46
+ | 🎚️ **Cost-aware tiers** | `economy` / `standard` / `frontier` model tiers with cost caps + context floors. Route cheap work to cheap models, escalate when it matters. Live cost $ + context % per run. |
47
+ | 🚦 **Quality gates** | `verification-before-completion`, `completeness-check`, `gate`, `verify` — built-in. Register your own. Composite helpers: `judgePanel`, `loopUntilDry`, `retry`. |
48
+ | 🧩 **Workflows-as-code** | Author multi-phase workflows in a JS DSL with `agent()`, `pipeline()`, `phase()`, `checkpoint()`. 5 builtins ship: adversarial-review, code-review, codebase-audit, deep-research, multi-perspective. Journaled + resumable. |
49
+ | 🖥️ **Fleet TUI** | `/fleet` opens an interactive panel: Runs, Tiers, Lifecycle, Workflows, Conversation viewer. Live widget, mid-run Steer/Stop, edit-resume, save-as. |
50
+ | ⏱️ **Scheduling** | Cron expressions, intervals, one-shot ISO datetimes. PID-locked scheduler, session-scoped (no catch-up). Background runs on isolated git worktrees. |
51
+ | 🔒 **Worktree isolation** | Background runs get isolated git worktrees (in-place fallback for non-git cwds). Foreground runs share the session cwd. |
52
+ | 📒 **RunLog + journaling** | Every run is journaled; interrupted workflows recover on restart. A results inbox lets the model pull completed background runs. |
53
+ | 🔄 **Edit-and-resume** | Re-run a workflow by replaying the unchanged prefix and re-running only the edited suffix. |
54
+ | 📡 **Vision built-in** | `describe_image` tool is wired into child sessions — agents can see screenshots and diagrams without leaving the fleet. |
55
+
56
+ ---
57
+
58
+ ## Quick start
59
+
60
+ ### Install
61
+
62
+ armory-fleet is a [pi extension](https://pi-coding-agent.dev/docs/extensions) — it loads inside pi, no build step.
63
+
64
+ ```bash
65
+ # 1. Add to your pi packages (~/local-dev/arsenal or your package dir)
66
+ pnpm add @getpipher/armory-fleet
67
+
68
+ # 2. Register in ~/.pi/agent/settings.json
69
+ ```
70
+
71
+ ```jsonc
72
+ // ~/.pi/agent/settings.json
73
+ {
74
+ "packages": [
75
+ "@getpipher/armory-fleet@0.12.0"
76
+ // + its armory siblings: armory-todo, armory-memory, vision, cursor
77
+ ]
78
+ }
79
+ ```
80
+
81
+ ```bash
82
+ # 3. Reload pi (/reload) and open the panel
83
+ pi
84
+ # inside pi → /fleet
85
+ ```
86
+
87
+ ### Your first subagent (model-callable tool)
88
+
89
+ The `subagent` tool is what the model calls to delegate a focused task. Every child is armory-native by default.
90
+
91
+ ```ts
92
+ // the agent calls this — not you
93
+ subagent({
94
+ agent: "general-purpose",
95
+ task: "Audit src/auth/ for token-handling bugs; report findings.",
96
+ // optional: model, lifecycle, todoId, background, isolation, schedule, maxTurns
97
+ });
98
+ ```
99
+
100
+ | Param | Effect |
101
+ |---|---|
102
+ | `agent` | Agent definition to spawn (from `agents/` or discovered). |
103
+ | `task` | The prompt handed to the child. |
104
+ | `model` | Override the session model. **Tip:** omit to inherit the session model, or use `Ollama/...` when the session is on Ollama — don't cross providers. |
105
+ | `lifecycle` | Run the task through a superpowers lifecycle (e.g. `default`) instead of a single delegate. |
106
+ | `todoId` | Link the run to an existing armory-todo entry. |
107
+ | `track` | Default `true` (syncs to armory-todo). Pass `false` only for throwaway lookups. |
108
+ | `background` | Fire without awaiting — run goes to the async pool on an isolated git worktree. |
109
+ | `isolation` | `worktree` (default for bg in a git repo) · `none` (in-place) · `auto`. |
110
+ | `schedule` | Cron (`0 9 * * 1-5`), interval (`30m`), or one-shot ISO datetime. Session-scoped, no catch-up. |
111
+ | `maxTurns` | Per-run turn budget (default 20). Raise for complex multi-step tasks. |
112
+
113
+ ### Your first workflow
114
+
115
+ Workflows are plain JS files evaluated in a sandboxed vm realm. The orchestration primitives — `agent`, `parallel`, `pipeline`, `phase`, `gate`, `judgePanel`, `loopUntilDry`, `retry`, `checkpoint`, `verify`, `workflow`, `log` — are **injected globals** (no imports). The only thing you `export` is `meta`.
116
+
117
+ ```js
118
+ // ship-feature.js — drop into a workflows/ dir discovered by WorkflowRegistry
119
+ export const meta = {
120
+ name: 'ship-feature',
121
+ description: 'Plan → implement → 3 parallel review angles with a gate',
122
+ phases: [{ title: 'Plan' }, { title: 'Implement' }, { title: 'Review' }],
123
+ }
124
+
125
+ phase('Plan')
126
+ const plan = await agent('Plan this feature: ' + args.task, { tier: 'low' })
127
+
128
+ phase('Implement')
129
+ const impl = await agent(`Implement the plan:\n${plan}`, { tier: 'medium' })
130
+
131
+ phase('Review')
132
+ const angles = ['security', 'performance', 'correctness']
133
+ const reviews = await parallel(
134
+ angles.map((a) => () => agent(`Review the implementation for ${a} issues.`, { tier: 'low' })),
135
+ )
136
+
137
+ // gate: revise the synthesis until it passes a validator
138
+ const synthesis = await gate(
139
+ async (_feedback, n) => n === 0
140
+ ? agent(`Synthesize ${reviews.length} reviews.`, { tier: 'low' })
141
+ : agent('Revise synthesis per feedback.', { tier: 'low' }),
142
+ (v) => typeof v === 'string' && v.length > 200 ? { ok: true } : { ok: false, feedback: 'more detail' },
143
+ { attempts: 3 },
144
+ )
145
+
146
+ return { plan, impl, reviews, synthesis }
147
+ ```
148
+
149
+ Open `/fleet → Workflows`, pick `ship-feature`, run it. The panel shows live phase progress; mid-run you can Steer (inject a message) or Stop. The realm also exposes `args`, `cwd`, and a `budget` object (`{ total, spent(), remaining() }`) so workflows can self-limit.
150
+
151
+ ### Your first lifecycle run
152
+
153
+ ```text
154
+ /fleet-implement Refactor the auth module to use the new session API --auto
155
+ ```
156
+
157
+ Runs `brainstorm → plan → implement → review → finish` autonomously. Drop `--auto` for checkpointed mode (pauses at each checkpoint; continue/revise/abort from `/fleet → Lifecycle`).
158
+
159
+ ---
160
+
161
+ ## Architecture
162
+
163
+ ```
164
+ ┌─────────────────────────────────────────────┐
165
+ │ pi host session │
166
+ │ (loads @getpipher/armory-fleet extension) │
167
+ └───────────────────────┬─────────────────────┘
168
+
169
+ ┌───────────────────────────────────────┼───────────────────────────────────────┐
170
+ ▼ ▼ ▼
171
+ subagent tool fleet tool /fleet panel
172
+ (model-callable) (workflow runner) (FleetView TUI)
173
+ │ │ │
174
+ ▼ ▼ ▼
175
+ createAgentSession() WorkflowController Runs · Tiers · Lifecycle
176
+ (pi SDK child) + ConcurrencyPool Workflows · Conversation
177
+ │ + adapters + live widget
178
+ ├─→ armory-todo sync + journal/resume
179
+ ├─→ armory-memory hydrate │
180
+ ├─→ vision (describe_image) ▼
181
+ ├─→ lifecycle + gates backend registry
182
+ └─→ tier routing (pi | Claude Code)
183
+ ```
184
+
185
+ ### Core engine
186
+
187
+ - **Engine primitive:** `createAgentSession()` from the pi SDK — child Pi sessions, in-memory or file-backed `SessionManager`, `ResourceLoader`. Each child is wrapped to emit `session_init` on subscribe so the fleet can track it from the first event.
188
+ - **Child loader:** `buildChildLoader()` threads armory-todo, armory-memory, and vision into every child's resource + tool set — armory-native from birth, cwd-agnostic.
189
+ - **Concurrency:** a single-slot lock for foreground runs + a `ConcurrencyPool` for parallel workflow branches.
190
+ - **Turn budget:** `engine/turn-budget.ts` caps each child's run; the `subagent` tool surfaces exhaustion as a structured status (not a silent truncation).
191
+
192
+ ### Armory integration (the uncopyable layer)
193
+
194
+ | Sibling | What the fleet wires in | Where |
195
+ |---|---|---|
196
+ | [armory-todo](https://github.com/getpipher/armory-todo) | Every run syncs to the cross-session TODO store. Pass `todoId` to link. | `src/todo-sync/` |
197
+ | [armory-memory](https://github.com/getpipher/armory-memory) | Children hydrate project memory on spawn. Shared port, cwd-agnostic. | `src/memory-hydrate/` |
198
+ | [vision](https://github.com/getpipher/vision) | `describe_image` tool is wired into child sessions. | `src/vision/` |
199
+ | [cursor](https://github.com/getpipher/cursor) | Children edit through the cursor extension when present. | (via child loader) |
200
+
201
+ ### Cross-harness backends
202
+
203
+ `src/backend/` ships a backend registry with **pi** (default) and **Claude Code** as peer backends. `detectClaude()` auto-discovers Claude; `PI_HOOK_PARITY` / `CLAUDE_HOOK_PARITY` tables keep both backends on equal footing. `hook-parity.ts` normalizes lifecycle/event hooks across harnesses. A `ResumeStore` persists backend session IDs so cross-harness runs can resume.
204
+
205
+ ### Superpowers lifecycle
10
206
 
11
- 1. **Armory-native integration** no external package can integrate with the armory suite; they don't own it. Agents that sync to armory-todo, hydrate from armory-memory, see via vision, and edit via cursor *by default* are uncopyable.
12
- 2. **Cross-harness peers** — only one attempt (kky42/pi-flow, 66⭐) runs Claude Code + Pi as peer subagent backends, and it's early. A dual-arsenal topology deserves a first-class implementation.
13
- 3. **Superpowers-native lifecycle** — only one attempt (teelicht/pi-superagents, 54⭐) wraps the superpowers skill pipeline, and it's synchronous-only.
207
+ The default lifecycle (`src/lifecycle/default.ts`) is the superpowers-native 5-phase pipeline:
14
208
 
15
- `armory-fleet` owns all three, then reaches parity on the rest (fleet TUI, scheduling, worktree isolation, cost accounting, quality patterns, workflows-as-code, event-bus/RPC) — to be the best pi-subagent package in the ecosystem.
209
+ | Phase | Skills loaded | Checkpoint? | Gates |
210
+ |---|---|---|---|
211
+ | `brainstorm` | `brainstorming` | ✅ | — |
212
+ | `plan` | `writing-plans` | ✅ | `completenessCheck` |
213
+ | `implement` | `executing-plans`, `test-driven-development`, `verification-before-completion` | ❌ | `verification-before-completion`, `completenessCheck`, `gate` |
214
+ | `review` | `requesting-code-review`, `receiving-code-review` | ✅ | — |
215
+ | `finish` | `finishing-a-development-branch` | — | — |
16
216
 
17
- ## Status
217
+ Custom lifecycles: drop a YAML file in your `lifecycles/` dir, register via `discoverLifecycles()`. Gates are registered on a `GateRegistry` (`fleet-register-gate` command for runtime extensibility).
18
218
 
19
- 🚧 **Pre-implementation.** The master PRD + 7-SPEC roadmap are written and self-reviewed; SPEC-1 (core engine + armory-todo sync) is the next brainstorm cycle.
219
+ ### Quality gates
20
220
 
21
- - **Master PRD:** [`PRD.md`](./PRD.md)
22
- - **Landscape research:** [`research/`](./research) (11+ packages mapped, 5 contenders deep-read)
23
- - **Roadmap:** PRD → SPEC-1 (core+todo-sync) SPEC-2 (deep armory) → SPEC-3 (cross-harness) → SPEC-4 (superpowers lifecycle) → SPEC-5a (ops) → SPEC-5b (fleet TUI) → SPEC-6 (power-user → v1.0)
221
+ Built-in (`src/lifecycle/gates/`): `verification-before-completion`, `completeness-check`, `gate`, `verify`.
222
+
223
+ Composite helpers (`src/workflows/helpers/`) usable from any workflow:
224
+
225
+ | Helper | What it does |
226
+ |---|---|
227
+ | `judgePanel` | Run N judge agents; majority/weighted verdict. |
228
+ | `loopUntilDry` | Re-run an agent until a dry-run gate passes. |
229
+ | `retry` | Retry an agent with backoff on failure. |
230
+ | `checkpoint` | Pause a workflow for human review. |
231
+ | `completeness-check` / `gate` / `verify` | Gate wrappers for workflow use. |
232
+
233
+ ### Cost-aware tiers
234
+
235
+ `src/tiers/` ships three built-in tiers:
236
+
237
+ | Tier | Models | Cost cap | Context floor |
238
+ |---|---|---|---|
239
+ | `economy` | `Ollama/minimax-m3:cloud` | — | — |
240
+ | `standard` | `Ollama/glm-5.2:cloud`, `Ollama/minimax-m3:cloud` | — | — |
241
+ | `frontier` | `anthropic/claude-sonnet-4`, `Ollama/glm-5.2:cloud` | $5 | 200k ctx |
242
+
243
+ Live cost $ and context % are tracked per run and surfaced in the Tiers view. Override per-run with `model`, or let the tier registry route based on the task class.
244
+
245
+ ### Operational runtime
246
+
247
+ `src/runtime/` — the async/scheduling spine:
248
+
249
+ - `async-runner.ts` — background dispatch (fire-and-forget).
250
+ - `run-journal.ts` + `run-log.ts` — durable run records; `reconcile.ts` reattaches orphaned runs on restart.
251
+ - `concurrency-pool.ts` — bounded parallel branches.
252
+ - `results-inbox.ts` — the model pulls completed background runs via the `fleet_results` tool.
253
+ - `resume.ts` — scan for resumable runs + workflows.
254
+
255
+ ### Scheduling + worktree
256
+
257
+ `src/scheduling/` — cron expressions (`expressions.ts`), a `Scheduler` with PID-locking (`pid-lock.ts`), session-scoped (no catch-up). `src/worktree/` — `WorktreeService` for isolated bg-run worktrees + `DiffService` for reviewable diffs.
258
+
259
+ ---
260
+
261
+ ## The fleet panel
262
+
263
+ `/fleet` opens an interactive TUI panel (TUI-only; in non-interactive modes use the `subagent` tool).
264
+
265
+ | View | What it shows |
266
+ |---|---|
267
+ | **Runs** | Running + recent subagents; status, cost, context %, agent, model. Action submenu: Steer, Stop, View conversation. |
268
+ | **Tiers** | Per-tier model lists, cost caps, context floors. Configure routing. |
269
+ | **Lifecycle** | Active lifecycle runs; Continue/Revise/Abort at checkpoints. |
270
+ | **Workflows** | Registered workflows + live runs. Run, edit-resume, save-as, view result, checkpoint. |
271
+ | **Conversation** | The full message timeline for any selected run. |
272
+
273
+ A live `FleetWidget` can render in the pi footer/overlay for at-a-glance fleet status while you work.
274
+
275
+ ### Slash commands
276
+
277
+ | Command | Purpose |
278
+ |---|---|
279
+ | `/fleet` | Open the interactive fleet panel (TUI). |
280
+ | `/fleet-implement <task> [--lifecycle <name>] [--auto]` | Run a task through the superpowers lifecycle. |
281
+ | `/fleet-register-gate` | Register a custom gate on the fleet gate registry (extensibility). |
282
+
283
+ ### Model-callable tools
284
+
285
+ | Tool | Purpose |
286
+ |---|---|
287
+ | `subagent` | Delegate a focused task to a child agent (sync foreground or async background). |
288
+ | `fleet` | Run + control fleet workflows (JS orchestration: `agent`, `pipeline`, `phase`, checkpoints). |
289
+ | `fleet_results` | Pull completed background run results from the inbox. |
290
+
291
+ ---
292
+
293
+ ## Workflows-as-code
294
+
295
+ Workflows are authored in a JS DSL (`src/workflows/source.ts` parses; `vm-realm.ts` evaluates). 5 builtins ship in `src/workflows/builtin/`:
296
+
297
+ | Workflow | Description | Phases |
298
+ |---|---|---|
299
+ | `adversarial-review` | Red-team + blue-team review with judge panel | Attack → Defend → Judge |
300
+ | `code-review` | 7 parallel review angles plus verification | Review → Verify |
301
+ | `codebase-audit` | File-tree scan with completeness check | Scan → Audit |
302
+ | `deep-research` | 3-round discovery loop with de-duplication | Discover → Synthesize |
303
+ | `multi-perspective` | 4 personas review the same artifact | Review → Merge |
304
+
305
+ Every workflow run is **journaled** (`workflows/journal.ts`) and **resumable**. `edit-resume` replays the unchanged prefix from cache and re-runs only the edited suffix. `runtime/controller.ts` orchestrates; `runtime/pause-gate.ts` handles checkpoints; `runtime/adapters.ts` binds the controller to the fleet's spawn + accounting.
306
+
307
+ ---
308
+
309
+ ## Roadmap
310
+
311
+ armory-fleet follows a PRD → SPEC-N (brainstorm → spec → plan → implementation) pipeline. **16/16 phases done through v0.12.0.**
312
+
313
+ | SPEC | Headline | Status | Artifact |
314
+ |---|---|---|---|
315
+ | PRD | Master PRD | ✅ done | `PRD.md` |
316
+ | RESEARCH | Landscape research (11+ packages, 5 deep-reads) | ✅ done | `research/` |
317
+ | SPEC-1 | Core engine + armory-todo sync | ✅ done | PR #1 · `547319b` |
318
+ | SPEC-2 | Deep armory integration (memory/vision/cursor) | ✅ done · @0.2.0 | PR #2 · `c6e727c` |
319
+ | SPEC-3 | Cross-harness peers (pi + Claude Code) | ✅ done · @0.3.0 | PR #4 · `5bb75fb` |
320
+ | SPEC-4 | Superpowers-native lifecycle | ✅ done · @0.4.0 | PR #5 · `67ff9b4` |
321
+ | SPEC-5a | Operational runtime (async/scheduling/worktree) | ✅ done · @0.5.2 | PR #6 · `52e3477` |
322
+ | SPEC-5b-1 | RunLog seam + Runs view | ✅ done · @0.6.0 | PR #7 · `54b1b10` |
323
+ | SPEC-5b-2 | Live widget + FleetView + Q9 | ✅ done · @0.7.0 | PR #8 · `9266a7` |
324
+ | SPEC-5b-3 | Conversation viewer + timeline fix | ✅ done · @0.8.0 | PR #9 · `adc0034` |
325
+ | SPEC-5b-4 | Mid-run steering (Steer) + Stop | ✅ done · @0.9.1 | PR #10 + #11 + #12 |
326
+ | SPEC-6-1 | Cost-aware tiers + cost $ + context % + Tiers view | ✅ done · @0.10.x | PR #15/#16/#17 |
327
+ | SPEC-6-2 | Quality gates + lifecycle hooks | ✅ done · @0.11.0 | PR #18 · `cda5e2b` |
328
+ | v0.11.1 | bg dispatch isolation split (non-git cwd fix) | ✅ done · @0.11.1 | PR #19 · `51956e0` |
329
+ | **SPEC-6-3** | **Workflows-as-code (release-gate completion)** | ✅ done · @0.12.0 | PR #21 · `9986ad1` |
330
+ | SPEC-6-4 | Event-bus RPC + live conversation viewer → v1.0 | 🚧 next | — |
331
+
332
+ See the [full release history](https://github.com/getpipher/armory-fleet/releases) and the [PRD](./PRD.md) §8 for the roadmap rationale.
333
+
334
+ ---
335
+
336
+ ## Ecosystem
337
+
338
+ armory-fleet is the orchestrator in the [getpipher](https://github.com/getpipher) armory suite — the default substrate it runs agents on:
339
+
340
+ | Package | Role |
341
+ |---|---|
342
+ | [armory-todo](https://github.com/getpipher/armory-todo) | Global cross-session TODO store (the fleet syncs every run to it). |
343
+ | [armory-memory](https://github.com/getpipher/armory-memory) | Project memory hydration for child agents. |
344
+ | [vision](https://github.com/getpipher/vision) | The `describe_image` tool, wired into fleet children. |
345
+ | [cursor](https://github.com/getpipher/cursor) | Custom editor component for the pi TUI. |
346
+
347
+ ---
348
+
349
+ ## Conventions
350
+
351
+ - **No build step** — extensions ship raw `.ts` via tsx at pi runtime. `pnpm typecheck` + `pnpm test:run` before release.
352
+ - **Tests** — `node:test` via tsx in `test/*.test.mts`, importing from `../src/...`. 593 passing.
353
+ - **Publish** — CI on `v*` tags using the getpipher `NPM_TOKEN` org secret (`release.yml` mirrors armory-todo: idempotent npm publish + GitHub Release).
354
+ - **Interactive-first UX** — every capability lands as a `/fleet` panel tab/view + action submenu first, then the model-callable tool action.
355
+
356
+ <a id="testing"></a>
357
+ ### Verify locally
358
+
359
+ ```bash
360
+ pnpm install
361
+ pnpm typecheck
362
+ pnpm test:run --test-timeout=30000 # 593/593
363
+ ```
364
+
365
+ ### Release-gate smoke (mandatory before any release)
366
+
367
+ ```bash
368
+ pi --no-extensions -e ./src/index.ts --no-session --approve
369
+ # inside: /fleet → Workflows → verify the 5 builtins render + a workflow runs end-to-end
370
+ ```
371
+
372
+ ---
24
373
 
25
374
  ## Compatibility
26
375
 
27
- pi `^0.81.1`.
376
+ - pi `^0.81.1`
377
+ - Node `>=22` (tsx runtime)
378
+ - Platform: macOS, Linux, WSL
28
379
 
29
380
  ## License
30
381
 
31
- MIT — see [LICENSE](./LICENSE).
382
+ MIT — see [LICENSE](./LICENSE). © RECTOR ([@rz1989s](https://github.com/rz1989s)).
383
+
384
+ <div align="center">
385
+
386
+ Built with Ihsan · Maintained by [RECTOR](https://github.com/rz1989s) · [getpipher](https://github.com/getpipher)
387
+
388
+ </div>
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@getpipher/armory-fleet",
3
- "version": "0.12.0",
3
+ "version": "0.12.2",
4
4
  "private": false,
5
5
  "description": "The armory suite's subagent orchestrator for the pi coding agent — a cross-harness, superpowers-native fleet where every agent is armory-native from birth.",
6
6
  "license": "MIT",
@@ -7,6 +7,30 @@ import { dirname } from "node:path";
7
7
  import type { AgentDef } from "../registry/frontmatter.ts";
8
8
  import type { MemoryHydratePort } from "../memory-hydrate/port.ts";
9
9
 
10
+ /** Shape of the pi resource-loader's current-skills context (the `cur` arg of skillsOverride).
11
+ * Generic over the skill + diagnostics types so the pi SDK's full `Skill`/`ResourceDiagnostic`
12
+ * types flow through unchanged. */
13
+ export interface InstalledSkills<S = { name: string }, D = unknown> {
14
+ skills: S[];
15
+ diagnostics: D;
16
+ }
17
+
18
+ /** #32: resolve the child's skill bundle from the agent's declared skills + the installed arsenal.
19
+ *
20
+ * - Agent declares specific skills → load only those (filtered from the installed arsenal).
21
+ * - Agent declares NO skills → load NO skills (lean substrate). This is the #32 fix: previously
22
+ * an agent with no `skills` field loaded ALL installed skills (~42 → ~570K substrate on every
23
+ * `general-purpose` dispatch, ~59% of a 976K context window from turn 1). Callers opt in to
24
+ * skills via the `subagent` tool's `skills` param (threaded as `skillsOverride`, which clones
25
+ * `agent.skills` before this runs).
26
+ * - Returning `[]` (not `cur.skills`) is the deliberate behavior change. */
27
+ export function resolveChildSkills<S extends { name: string }, D>(agent: Pick<AgentDef, "skills">, cur: InstalledSkills<S, D>): InstalledSkills<S, D> {
28
+ const selected = agent.skills && agent.skills.length
29
+ ? cur.skills.filter((s) => agent.skills!.includes(s.name))
30
+ : [];
31
+ return { skills: selected, diagnostics: cur.diagnostics };
32
+ }
33
+
10
34
  /** Fixed pseudo-cwd for the global cross-project user memory scope. */
11
35
  export const USER_PSEUDO_CWD = "/__armory-fleet-user__";
12
36
 
@@ -36,12 +60,6 @@ export function buildChildLoader(opts: ChildLoaderOpts): DefaultResourceLoader {
36
60
  noExtensions: true,
37
61
  systemPromptOverride: (base) =>
38
62
  composeChildPrompt({ rolePrompt: opts.agent.rolePrompt, memoryBlock, base: base ?? "" }),
39
- skillsOverride: (cur) => ({
40
- skills:
41
- opts.agent.skills && opts.agent.skills.length
42
- ? cur.skills.filter((s) => opts.agent.skills!.includes(s.name))
43
- : cur.skills,
44
- diagnostics: cur.diagnostics,
45
- }),
63
+ skillsOverride: (cur) => resolveChildSkills(opts.agent, cur),
46
64
  });
47
- }
65
+ }
@@ -142,6 +142,11 @@ export interface SpawnOptions {
142
142
  modelRegistry?: ModelRegistryLike;
143
143
  /** SPEC-6-3: workflow adapter tier override — replaces agent.tier before model resolution. */
144
144
  tierOverride?: string;
145
+ /** #31: when true, the caller asserts this dispatch will NOT mutate the working directory
146
+ * (review/audit/research). Read-only dispatches bypass the foreground single-slot lock so
147
+ * multiple readOnly dispatches — and/or a readOnly dispatch alongside a write dispatch — can
148
+ * run in parallel. Mislabeling a write dispatch as readOnly risks in-place edit conflicts. */
149
+ readOnly?: boolean;
145
150
  }
146
151
 
147
152
  export interface SpawnResult {
@@ -158,6 +163,11 @@ export interface SpawnResult {
158
163
  /** SPEC-6-1: final context tokens (calcContextTokens of the last usage). */
159
164
  contextTokens?: number;
160
165
  error?: string;
166
+ /** #39: true when the failure is a retryable provider rate-limit / auth failure
167
+ * (stopReason "error"). The subagent tool retries once with the configured `modelFallback`
168
+ * when this is set. Non-retryable failures (turn budget, agent-not-found, EMPTY_RESULT,
169
+ * abort, lock busy) leave this unset. */
170
+ retryable?: boolean;
161
171
  }
162
172
 
163
173
  export async function spawnSubagent(opts: SpawnOptions): Promise<SpawnResult> {
@@ -165,15 +175,25 @@ export async function spawnSubagent(opts: SpawnOptions): Promise<SpawnResult> {
165
175
  const maxTurns = opts.maxTurns ?? DEFAULT_MAX_TURNS;
166
176
  const startedAt = Date.now();
167
177
 
168
- // concurrency=1 (SPEC-1 §9.2) peek before generating a runId so a rejected
169
- // call doesn't mint a discarded id; the held id is named in the message.
170
- const busyId = opts.lock.current();
171
- if (busyId !== null) {
172
- return fail("", startedAt, `a subagent is already running (concurrency=1 in v0.1); wait for ${busyId} to finish or abort it first`, opts.agent);
173
- }
174
- const runId = genRunId();
175
- if (!opts.lock.tryAcquire(runId)) {
176
- return fail("", startedAt, "concurrency lock unexpectedly unavailable", opts.agent);
178
+ // #31: read-only dispatches (review/audit/research) bypass the foreground single-slot lock
179
+ // the caller asserts no cwd mutation, so the in-place edit-conflict guard doesn't apply and
180
+ // multiple readOnly dispatches (and/or a readOnly alongside a write dispatch) may run in
181
+ // parallel. The lock is still acquired for write dispatches (default), preserving concurrency=1.
182
+ const readOnly = opts.readOnly ?? false;
183
+ let runId: string;
184
+ if (readOnly) {
185
+ runId = genRunId();
186
+ } else {
187
+ // concurrency=1 (SPEC-1 §9.2) — peek before generating a runId so a rejected
188
+ // call doesn't mint a discarded id; the held id is named in the message.
189
+ const busyId = opts.lock.current();
190
+ if (busyId !== null) {
191
+ return fail("", startedAt, `a subagent is already running (concurrency=1 in v0.1); wait for ${busyId} to finish or abort it first`, opts.agent);
192
+ }
193
+ runId = genRunId();
194
+ if (!opts.lock.tryAcquire(runId)) {
195
+ return fail("", startedAt, "concurrency lock unexpectedly unavailable", opts.agent);
196
+ }
177
197
  }
178
198
 
179
199
  try {
@@ -282,6 +302,10 @@ export async function spawnSubagent(opts: SpawnOptions): Promise<SpawnResult> {
282
302
  let costTotal = 0;
283
303
  let contextTokens = 0;
284
304
  let turnIdx = -1;
305
+ // #26/#22: declared before subscribe() because some child sessions emit events
306
+ // synchronously inside subscribe() (temporal-dead-zone guard).
307
+ let modelError: string | undefined; // model-call failure surfaced via stopReason "error"
308
+ let sawAssistantMessage = false; // #22: did the child emit any assistant message_end at all?
285
309
 
286
310
  const onSignalAbort = (): void => { aborted = true; void session.abort(); };
287
311
  opts.signal?.addEventListener("abort", onSignalAbort);
@@ -297,8 +321,18 @@ export async function spawnSubagent(opts: SpawnOptions): Promise<SpawnResult> {
297
321
  } else if (e.type === "turn_end") {
298
322
  if (budget.consume()) void session.abort();
299
323
  } else if (e.type === "message_end" && e.message?.role === "assistant") {
324
+ sawAssistantMessage = true;
300
325
  const text = e.message.content?.map((c) => (c.type === "text" ? c.text ?? "" : "")).join("") ?? "";
301
- if (text) finalText = text;
326
+ // #26/#22: a model-call failure (401, provider down, rate limit) surfaces as
327
+ // stopReason "error". The SDK retries internally; if it still ends with an error
328
+ // stopReason, capture it so the run is marked failed (not completed-with-empty) —
329
+ // the controller gets an actionable error instead of "(no tool output)".
330
+ const stopReason = (e.message as { stopReason?: string }).stopReason;
331
+ if (stopReason === "error") {
332
+ modelError = text || `model call ended with stopReason 'error' (provider/auth failure or rate limit) for model '${model}'`;
333
+ } else {
334
+ if (text) finalText = text;
335
+ }
302
336
  // SPEC-5b-2 (Q9): accumulate REAL tokens (input+output+cacheRead+cacheWrite), not cost.total (dollars).
303
337
  const u = e.message.usage;
304
338
  const turnTokens = (u?.input ?? 0) + (u?.output ?? 0) + (u?.cacheRead ?? 0) + (u?.cacheWrite ?? 0);
@@ -343,18 +377,53 @@ export async function spawnSubagent(opts: SpawnOptions): Promise<SpawnResult> {
343
377
  status = "aborted";
344
378
  error = tier?.costCap && costTotal > tier.costCap ? `budget_exceeded (cost $${costTotal.toFixed(4)} > cap $${tier.costCap})` : "aborted by user";
345
379
  } else if (budget.count() >= maxTurns) {
380
+ // #25: surface a coherent partial, not a mid-sentence 200-char cut. The controller reads
381
+ // `res.error` (the tool surfaces error, not finalText, for failed runs), so the partial must
382
+ // live here. 4000 chars (~600 tokens) is enough for any structured summary the model emitted;
383
+ // a truncation marker names the run log for the full output. (The wind-down nudge — injecting
384
+ // a "you have ~N turns left, emit your partial now" message before the hard cut — is a
385
+ // future enhancement tracked in #25; it needs mid-loop injection semantics.)
386
+ const PARTIAL_WINDOW = 4000;
387
+ const partial = finalText.length > PARTIAL_WINDOW
388
+ ? finalText.slice(0, PARTIAL_WINDOW) + "\n…(partial truncated — see run log for full output)"
389
+ : finalText;
346
390
  status = "failed";
347
- error = `hit turn budget (${maxTurns}) mid-task; partial result: ${finalText.slice(0, 200)}`;
391
+ error = `hit turn budget (${maxTurns}) mid-task; partial result:\n${partial}`;
348
392
  } else if (runError) {
349
393
  status = "failed";
350
394
  error = runError;
395
+ } else if (modelError) {
396
+ // #26: a 401/provider/rate-limit failure that the SDK surfaced via stopReason "error"
397
+ // after exhausting retries. Without this, the run fell through to `completed` with an
398
+ // empty finalText — the controller saw "(no tool output)" and couldn't tell a broken
399
+ // model from a no-op run.
400
+ //
401
+ // Precedence note (PR #30 review): a late error-stop overrides a prior successful turn.
402
+ // If turn 1 set finalText (valid output) and turn 2 hit stopReason "error", the run is
403
+ // marked failed with the error — the run IS incomplete, and the error is more actionable
404
+ // to the controller than a partial result. finalText is preserved (not cleared) so
405
+ // finishRun + the run log still carry the partial; only the surfaced status is failed.
406
+ // Gating this on `!finalText` (only fail if no prior output) is a future design call, not
407
+ // this fix — the current "last error wins" is the defensible default.
408
+ status = "failed";
409
+ error = modelError;
410
+ } else if (!sawAssistantMessage) {
411
+ // #22: prompt() resolved cleanly but the child produced NO assistant message_end at all.
412
+ // A real agent loop always emits at least one assistant message; zero means a silent
413
+ // failure (provider hung, empty response, premature exit). Treat as a structured
414
+ // EMPTY_RESULT so orchestration can escalate models or retry, rather than silently
415
+ // succeeding with empty output the controller can't distinguish from a no-op.
416
+ status = "failed";
417
+ error = `EMPTY_RESULT: child session produced no assistant output for model '${model}' (possible provider/auth failure, empty response, or premature exit)`;
351
418
  } else {
352
419
  status = "completed";
353
420
  }
354
421
 
355
- return await finishRun(opts, runId, startedAt, status, finalText, todoId, priorStatus, error, agentDef.name, model, tokenTotal, costTotal, contextTokens);
422
+ return await finishRun(opts, runId, startedAt, status, finalText, todoId, priorStatus, error, agentDef.name, model, tokenTotal, costTotal, contextTokens, modelError ? true : undefined);
356
423
  } finally {
357
- opts.lock.release();
424
+ // #31: a readOnly dispatch never acquired the lock — don't release what it didn't take
425
+ // (releasing a lock held by another concurrent write dispatch would corrupt serialization).
426
+ if (!readOnly) opts.lock.release();
358
427
  }
359
428
  }
360
429
 
@@ -372,6 +441,7 @@ async function finishRun(
372
441
  opts: SpawnOptions, runId: string, startedAt: number,
373
442
  status: FleetRunStatus, finalText: string, todoId: string | null, priorStatus: string | undefined,
374
443
  error: string | undefined, agentName: string, model: string, tokenTotal = 0, costTotal = 0, contextTokens = 0,
444
+ retryable?: boolean,
375
445
  ): Promise<SpawnResult> {
376
446
  if (finalizedRunIds.has(runId)) {
377
447
  // Already finalized — return the existing registry record's result without re-appending.
@@ -380,7 +450,7 @@ async function finishRun(
380
450
  status: existing?.status ?? status, finalText: existing?.resultSummary ?? finalText,
381
451
  runId, todoId, agent: agentName, model,
382
452
  durationMs: existing?.endedAt ? existing.endedAt - startedAt : Date.now() - startedAt,
383
- tokenTotal, costTotal, contextTokens, error,
453
+ tokenTotal, costTotal, contextTokens, error, retryable,
384
454
  };
385
455
  }
386
456
  finalizedRunIds.add(runId);
@@ -414,6 +484,6 @@ async function finishRun(
414
484
  }
415
485
  return {
416
486
  status, finalText, runId, todoId, agent: agentName, model,
417
- durationMs: endedAt - startedAt, tokenTotal, costTotal, contextTokens, error,
487
+ durationMs: endedAt - startedAt, tokenTotal, costTotal, contextTokens, error, retryable,
418
488
  };
419
489
  }
package/src/index.ts CHANGED
@@ -294,10 +294,15 @@ export default async function (pi: ExtensionAPI): Promise<void> {
294
294
  deps.runLog = new RunLog(join(dir, "conversations"));
295
295
  // v0.10.2: pass the in-memory RunRegistry so reconcile syncs it too — otherwise orphaned
296
296
  // (process-gone) runs keep status:"running" in memory and the live widget shows a stale ▶ forever.
297
- const reconciled = reconcileRuns(deps.runLog, { runRegistry: deps.runRegistry });
298
- if (reconciled.length > 0) {
299
- ctx.ui.notify(`reconciled ${reconciled.length} interrupted fleet run${reconciled.length > 1 ? "s" : ""} (marked aborted)`, "info");
300
- }
297
+ // #22 bg-watchdog: pass todoSync so a process-gone run's linked TODO is reverted to open
298
+ // (retryable) with a WORKER_EXITED_WITHOUT_RESULT note, not stuck in_progress forever.
299
+ // Fire-and-forget (async) so the asyncRunner setup below isn't blocked.
300
+ void reconcileRuns(deps.runLog, { runRegistry: deps.runRegistry, todoSync: deps.todoSync })
301
+ .then((reconciled) => {
302
+ if (reconciled.length > 0) {
303
+ ctx.ui.notify(`reconciled ${reconciled.length} interrupted fleet run${reconciled.length > 1 ? "s" : ""} (marked aborted; linked TODOs reverted to open)`, "info");
304
+ }
305
+ });
301
306
  deps.asyncRunner = {
302
307
  worktree: new WorktreeService({ rootDir: ctx.cwd }),
303
308
  diff: new DiffService(),
@@ -340,6 +345,7 @@ export default async function (pi: ExtensionAPI): Promise<void> {
340
345
  getModelContextWindow,
341
346
  cwd: ctx.cwd,
342
347
  runLog: deps.runLog,
348
+ todoSync: deps.todoSync, // #22 bg-watchdog: periodic probe reverts process-gone runs' TODOs
343
349
  });
344
350
  fleetWidget.start();
345
351
 
@@ -697,19 +697,31 @@ export class FleetPanel extends Container {
697
697
  }
698
698
  // SPEC-6-3: Workflows view — direct p/u/x controls + host-only intent completion
699
699
  if (this.view === "workflows") {
700
+ // #27: classify the key FIRST. Non-action keys (Down/Up/PageUp/PageDown) are forwarded
701
+ // to the list BEFORE the sel check, so nav still works when no row is selected (empty-list
702
+ // edge case the reviewer flagged — the prior `if (!sel) return` at the top swallowed nav
703
+ // keys the same way the original `if (!action) return` did).
704
+ const keyAction: Record<string, WorkflowPanelAction> = {
705
+ r: "run", e: "edit-resume", o: "open", p: "pause", u: "resume", x: "stop", s: "save", v: "view-result", c: "respond",
706
+ }
707
+ const action = keyAction[data]
708
+ if (!action) {
709
+ // Not a workflow action key — forward to the list so Down/Up/PageUp/PageDown move the
710
+ // selection cursor (#27). Without this, every non-action key was swallowed here and the
711
+ // bottom-of-handleInput `this.list.handleInput(data)` was never reached for Workflows,
712
+ // so the → cursor could never move off the first row (blocking all run-row actions).
713
+ this.list.handleInput(data)
714
+ this.invalidate()
715
+ return
716
+ }
700
717
  const sel = this.list.getSelectedItem()
701
- if (!sel) return
718
+ if (!sel) return // an action key was pressed but there's no row to act on
702
719
  const parsed = parseWorkflowPanelKey(sel.value)
703
720
  const item: WorkflowPanelItem = parsed.kind === "definition"
704
721
  ? { kind: "definition", definition: this.deps.workflowRegistry.get(parsed.name) ?? { name: parsed.name, description: "", phases: [], sourceText: "", body: "", executable: "", source: "builtin", filePath: "" } }
705
722
  : { kind: "run", run: this.deps.workflowStore.get(parsed.runId) ?? { runId: parsed.runId, name: parsed.runId, script: "", mode: "auto", status: "completed", startedAt: 0, currentPhase: "default", phases: [], childRunIds: [], logs: [], tokenTotal: 0, costTotal: 0 } }
706
723
  const available = actionsForWorkflowItem(item)
707
724
 
708
- const keyAction: Record<string, WorkflowPanelAction> = {
709
- r: "run", e: "edit-resume", o: "open", p: "pause", u: "resume", x: "stop", s: "save", v: "view-result", c: "respond",
710
- }
711
- const action = keyAction[data]
712
- if (!action) return
713
725
  if (!available.includes(action)) {
714
726
  this.onNotify(`action '${action}' not available for this item`, "warning")
715
727
  return
@@ -46,6 +46,9 @@ export interface FleetWidgetDeps {
46
46
  cwd?: string;
47
47
  /** SPEC-6-2: RunLog for the periodic liveness probe (reconcileRuns). Optional — absent → no periodic probe. */
48
48
  runLog?: import("../runtime/run-log.ts").RunLog;
49
+ /** #22 bg-watchdog: when wired, the periodic probe reverts a process-gone run's linked TODO to
50
+ * open (retryable) with a WORKER_EXITED_WITHOUT_RESULT note. Optional — absent → no TODO transition. */
51
+ todoSync?: import("../todo-sync/port.ts").TodoSyncPort;
49
52
  }
50
53
 
51
54
  export class FleetWidgetController {
@@ -69,9 +72,10 @@ export class FleetWidgetController {
69
72
  this.unsubs.push(this.deps.runRegistry.subscribe(() => this.render()));
70
73
  if (this.deps.bgRuns) this.unsubs.push(this.deps.bgRuns.subscribe(() => this.render()));
71
74
  // SPEC-6-2: periodic liveness probe — reconciles dead orphans every 60s.
75
+ // #22 bg-watchdog: pass todoSync so a process-gone run's TODO is reverted (not stuck in_progress).
72
76
  if (this.deps.runLog) {
73
77
  this.livenessTimerId = this.setIntervalFn(() => {
74
- reconcileRuns(this.deps.runLog!, { runRegistry: this.deps.runRegistry });
78
+ void reconcileRuns(this.deps.runLog!, { runRegistry: this.deps.runRegistry, todoSync: this.deps.todoSync }).catch(() => {});
75
79
  }, 60_000);
76
80
  (this.livenessTimerId as { unref?: () => void }).unref?.();
77
81
  }
@@ -4,6 +4,7 @@
4
4
  import { WorkflowJournal } from "../workflows/journal.ts";
5
5
  import type { RunLog } from "./run-log.ts";
6
6
  import type { RunRegistry, RunRecord } from "../engine/run-registry.ts";
7
+ import type { TodoSyncPort } from "../todo-sync/port.ts";
7
8
 
8
9
  export type Liveness = "alive" | "dead";
9
10
 
@@ -25,13 +26,20 @@ export interface ReconcileOpts {
25
26
  graceMs?: number;
26
27
  now?: number;
27
28
  runRegistry?: RunRegistry;
29
+ /** #22 bg-watchdog: when wired, a process-gone run's linked TODO is reverted to open
30
+ * (retryable) with a WORKER_EXITED_WITHOUT_RESULT note, so it doesn't stay in_progress
31
+ * forever after a worker exits without a terminal record. Best-effort (the run is already
32
+ * marked aborted in the log + registry). */
33
+ todoSync?: TodoSyncPort;
28
34
  }
29
35
 
30
- /** Returns the runIds it marked aborted. Probe-driven (SPEC-6-2); idempotent. */
31
- export function reconcileRuns(log: RunLog, opts: ReconcileOpts = {}): string[] {
36
+ /** Returns the runIds it marked aborted. Probe-driven (SPEC-6-2); idempotent.
37
+ * #22: async awaits the best-effort TODO transition for each aborted run with a todoId. */
38
+ export async function reconcileRuns(log: RunLog, opts: ReconcileOpts = {}): Promise<string[]> {
32
39
  const grace = opts.graceMs ?? 60_000;
33
40
  const now = opts.now ?? Date.now();
34
41
  const reg = opts.runRegistry;
42
+ const todoSync = opts.todoSync;
35
43
  const aborted: string[] = [];
36
44
  for (const meta of log.scanMeta()) {
37
45
  if (meta.status !== "running") continue;
@@ -44,6 +52,16 @@ export function reconcileRuns(log: RunLog, opts: ReconcileOpts = {}): string[] {
44
52
  endedAt: now, resultSummary: "process-gone (probe)", tokenTotal: meta.tokenTotal,
45
53
  });
46
54
  reg?.update(meta.runId, { status: "aborted", endedAt: now });
55
+ // #22 bg-watchdog: transition the linked TODO so a worker that exited without a terminal
56
+ // record doesn't leave its fleet TODO stuck in_progress forever. markRunTodoReverted with
57
+ // priorStatus=undefined reverts a fleet-created TODO to open (retryable) + appends the
58
+ // reason note. (The link path only accepts open/in_progress TODOs, so reverting to open is
59
+ // the correct recovery for a linked one too — its prior was open/in_progress.)
60
+ if (meta.todoId && todoSync) {
61
+ try {
62
+ await todoSync.markRunTodoReverted(meta.todoId, undefined, "WORKER_EXITED_WITHOUT_RESULT: process gone (probe)");
63
+ } catch { /* best-effort: the run is already marked aborted in the log + registry */ }
64
+ }
47
65
  aborted.push(meta.runId);
48
66
  }
49
67
  return aborted;
@@ -12,11 +12,30 @@ const FLEET_SOURCE = "armory-fleet";
12
12
  const FLEET_TAG = "fleet-run";
13
13
  const OPEN_STATES: Status[] = ["open", "in_progress"];
14
14
 
15
+ /** #34: cap the task excerpt written into the TODO notes well below armory-todo's maxNotesBytes
16
+ * (8192). The full task already lives in the run-log + the fleet run record; the TODO is a
17
+ * tracking stub, not a transcript. 1024 leaves ~7KB headroom for the fleet-run header + the
18
+ * progress/done/reverted appends during the run. */
19
+ const TASK_EXCERPT_CAP = 1024;
20
+ /** Cap appended note lines (done-result / reverted-reason). The #25 turn-budget partial can be
21
+ * ~4000 chars; appending it verbatim would accumulate toward the cap on retries. */
22
+ const NOTE_LINE_CAP = 500;
23
+
15
24
  function titleFor(run: RunMeta): string {
16
25
  const raw = `[${run.agent}] ${run.task}`.trim();
17
26
  return raw.length > 120 ? raw.slice(0, 117) + "…" : raw;
18
27
  }
19
28
 
29
+ function taskExcerptFor(task: string): string {
30
+ return task.length > TASK_EXCERPT_CAP
31
+ ? task.slice(0, TASK_EXCERPT_CAP) + "…[truncated; full task in run-log]"
32
+ : task;
33
+ }
34
+
35
+ function capNoteLine(s: string): string {
36
+ return s.length > NOTE_LINE_CAP ? s.slice(0, NOTE_LINE_CAP) + "…" : s;
37
+ }
38
+
20
39
  /** Append a note line to a todo (read-then-write; updateTodo replaces notes). */
21
40
  function appendNote(id: string, line: string): void {
22
41
  const t = getTodo(id);
@@ -54,7 +73,7 @@ export class ArmoryTodoAdapter implements TodoSyncPort {
54
73
  source: FLEET_SOURCE,
55
74
  priority: "med",
56
75
  tags: [FLEET_TAG],
57
- notes: `fleet-run:${run.runId}\n\nTask: ${run.task}`,
76
+ notes: `fleet-run:${run.runId}\n\nTask: ${taskExcerptFor(run.task)}`,
58
77
  });
59
78
  updateTodo(created.id, { status: "in_progress" });
60
79
  return { todoId: created.id }; // priorStatus undefined -> created
@@ -69,7 +88,7 @@ export class ArmoryTodoAdapter implements TodoSyncPort {
69
88
  // linked -> restore prior (user owns the close)
70
89
  updateTodo(todoId, { status: priorStatus as Status });
71
90
  }
72
- appendNote(todoId, `fleet-run done: ${result}`);
91
+ appendNote(todoId, `fleet-run done: ${capNoteLine(result)}`);
73
92
  }
74
93
 
75
94
  async markRunTodoReverted(todoId: string | null, priorStatus: string | undefined, reason: string): Promise<void> {
@@ -79,7 +98,7 @@ export class ArmoryTodoAdapter implements TodoSyncPort {
79
98
  } else {
80
99
  updateTodo(todoId, { status: priorStatus as Status });
81
100
  }
82
- appendNote(todoId, `fleet-run reverted: ${reason}`);
101
+ appendNote(todoId, `fleet-run reverted: ${capNoteLine(reason)}`);
83
102
  }
84
103
 
85
104
  async updateLifecycleProgress(todoId: string, progressBlock: string): Promise<void> {
@@ -29,10 +29,21 @@ export const subagentParams = Type.Object({
29
29
  ], { description: "Edit isolation for background runs. 'worktree' = git worktree (requires a git repo; fails sync if not). 'none' = in-place in cwd (no isolation; parallel edits may conflict). 'auto' (default) = worktree when cwd is a git repo, in-place otherwise." })),
30
30
  schedule: Type.Optional(Type.String({ description: 'Schedule the run instead of firing now: a cron string ("0 9 * * 1-5"), an interval ("30m"/"2h"), or a one-shot ISO datetime ("2026-07-25T14:00"). Returns { scheduleId, nextFire }. Session-scoped (fires only while pi is open); no catch-up.' })),
31
31
  maxTurns: Type.Optional(Type.Number({ description: 'Per-run turn budget (default 20). Raise for complex multi-step tasks (e.g. 40) so the subagent doesn\'t hit the budget mid-task; lower for trivial lookups.' })),
32
+ readOnly: Type.Optional(Type.Boolean({ description: 'Default false. Pass true ONLY for dispatches that will NOT mutate the working directory (review/audit, or research that writes no scratch files). A readOnly dispatch bypasses the foreground single-slot lock so multiple readOnly dispatches — and/or a readOnly alongside a write dispatch — can run in parallel. The caller is responsible for the assertion: mislabeling a dispatch that edits as readOnly risks in-place edit conflicts. Has no effect on background/scheduled runs (they use their own locks).' })),
33
+ skills: Type.Optional(Type.Array(Type.String(), { description: 'Skills to load for this dispatch (opt-in). By default a dispatch loads NO skills (#32 — lean substrate; previously an agent with no skills field loaded ALL ~42 installed skills, ~570K tokens / ~59% of context). Pass skill names from the installed arsenal (e.g. ["executing-plans", "test-driven-development"]) to opt in. For a direct dispatch, this replaces the agent\'s frontmatter skills (pass [] to load zero). For a lifecycle dispatch, this is ADDITIVE — the phase\'s designed skill bundle always loads and these are merged on top (a caller cannot strip a phase\'s required skills).' })),
34
+ modelFallback: Type.Optional(Type.String({ description: 'Model to retry with if the primary dispatch fails with a retryable provider rate-limit / auth failure (stopReason "error"). The fleet retries ONCE on this model and relinks the same tracked todo. Surface the model that served the retry in the result details (retriedWithModel). Per the AGENTS.md "Ollama primary + OpenRouter fallback" pattern. No effect on non-retryable failures (turn budget, agent-not-found, abort). Direct foreground dispatches only — background/scheduled/lifecycle retries are a follow-up.' })),
32
35
  });
33
36
 
34
37
  export type SubagentInput = Static<typeof subagentParams>;
35
38
 
39
+ /** #32: merge a lifecycle phase's skill bundle with the caller's `skills` param.
40
+ * Additive + deduped — the phase's designed skills always load; the caller can add extras but
41
+ * cannot strip phase skills (avoids the footgun where a caller passing `skills: ["tdd"]` with
42
+ * `lifecycle: "default"` would silently drop `brainstorming` from the brainstorm phase). */
43
+ export function mergeLifecycleSkills(phaseSkills: string[] | undefined, callerSkills: string[] | undefined): string[] {
44
+ return [...new Set([...(phaseSkills ?? []), ...(callerSkills ?? [])])];
45
+ }
46
+
36
47
  export interface SubagentToolDeps {
37
48
  registry: Map<string, AgentDef>;
38
49
  runRegistry: RunRegistry;
@@ -76,6 +87,9 @@ export function createSubagentTool(deps: SubagentToolDeps) {
76
87
  "Use subagent to delegate an isolated, well-scoped task to a named agent; it runs in the foreground and returns the result + a runId.",
77
88
  "Pass todoId to link the run to an existing open todo you see in the Open TODOs block; otherwise fleet creates a tracked fleet task.",
78
89
  "Pass track:false only for trivial throwaway lookups that don't represent real work.",
90
+ "Pass readOnly:true for dispatches that will NOT edit the working directory (review/audit, or research that writes no scratch files). It bypasses the foreground single-slot lock so multiple readOnly dispatches can run in parallel. Only use it when you are certain the child won't mutate cwd — mislabeling risks edit conflicts.",
91
+ "By default a dispatch loads NO skills (lean substrate). If the task needs a skill (e.g. test-driven-development for a TDD task, executing-plans for a plan-execution task), pass its name in the `skills` array to opt in — loading all skills by default wastes ~59% of the context window.",
92
+ "Pass `modelFallback` so a transient provider rate-limit / auth failure (stopReason 'error') auto-retries once on the fallback model instead of failing the dispatch. Per the AGENTS.md 'Ollama primary + OpenRouter fallback' pattern — don't let infra limits break a dispatch chain.",
79
93
  ],
80
94
  parameters: subagentParams,
81
95
  async execute(_toolCallId: string, params: SubagentInput, signal: AbortSignal, _onUpdate: unknown, _ctx: any) {
@@ -101,11 +115,12 @@ export function createSubagentTool(deps: SubagentToolDeps) {
101
115
  ...deps.lifecycleDeps,
102
116
  spawn: async (o) => spawnSubagent({
103
117
  agent: o.agent, task: o.task, lifecycleTodoId: o.lifecycleTodoId, model: o.model,
104
- skillsOverride: o.skills, backendOverride: o.backend,
118
+ skillsOverride: mergeLifecycleSkills(o.skills, params.skills), backendOverride: o.backend,
105
119
  registry: deps.registry, todoSync: deps.todoSync, runRegistry: deps.runRegistry, lock: deps.lock,
106
120
  backendRegistry: deps.backendRegistry, parentModel: deps.parentModel, parentCwd: deps.parentCwd, runLog: deps.runLog, signal,
107
121
  maxTurns: params.maxTurns,
108
122
  tierRegistry: deps.tierRegistry, modelRegistry: deps.modelRegistry,
123
+ readOnly: params.readOnly,
109
124
  }),
110
125
  };
111
126
  const res = await runLifecycle(params.task, params.lifecycle, {
@@ -127,6 +142,8 @@ export function createSubagentTool(deps: SubagentToolDeps) {
127
142
  todoId: params.todoId,
128
143
  track: params.track,
129
144
  model: params.model,
145
+ readOnly: params.readOnly,
146
+ skillsOverride: params.skills,
130
147
  registry: deps.registry,
131
148
  todoSync: deps.todoSync,
132
149
  runRegistry: deps.runRegistry,
@@ -139,15 +156,49 @@ export function createSubagentTool(deps: SubagentToolDeps) {
139
156
  maxTurns: params.maxTurns,
140
157
  tierRegistry: deps.tierRegistry, modelRegistry: deps.modelRegistry,
141
158
  });
142
- const isError = res.status === "failed" || res.status === "aborted";
159
+ // #39: auto-retry on a retryable provider rate-limit / auth failure (stopReason "error").
160
+ // The primary run reverted its linked todo to open (finishRun -> markRunTodoReverted), so the
161
+ // retry relinks the SAME todoId to continue the tracked task. Retry ONCE, only on the direct
162
+ // foreground path, only if a distinct fallback model was provided. The retry's runId differs
163
+ // from the primary's (each spawnSubagent call mints its own); details.retriedWithModel marks it.
164
+ let retriedWithModel: string | undefined;
165
+ let finalRes = res;
166
+ if (
167
+ res.status === "failed" && res.retryable && params.modelFallback &&
168
+ params.modelFallback !== res.model && !signal.aborted
169
+ ) {
170
+ finalRes = await spawnSubagent({
171
+ agent: params.agent,
172
+ task: params.task,
173
+ todoId: res.todoId ?? undefined,
174
+ track: params.track,
175
+ model: params.modelFallback,
176
+ readOnly: params.readOnly,
177
+ skillsOverride: params.skills,
178
+ registry: deps.registry,
179
+ todoSync: deps.todoSync,
180
+ runRegistry: deps.runRegistry,
181
+ lock: deps.lock,
182
+ backendRegistry: deps.backendRegistry,
183
+ parentModel: deps.parentModel,
184
+ parentCwd: deps.parentCwd,
185
+ runLog: deps.runLog,
186
+ signal,
187
+ maxTurns: params.maxTurns,
188
+ tierRegistry: deps.tierRegistry, modelRegistry: deps.modelRegistry,
189
+ });
190
+ retriedWithModel = params.modelFallback;
191
+ }
192
+ const isError = finalRes.status === "failed" || finalRes.status === "aborted";
143
193
  return {
144
- content: [{ type: "text" as const, text: isError ? (res.error ?? res.status) : res.finalText }],
194
+ content: [{ type: "text" as const, text: isError ? (finalRes.error ?? finalRes.status) : finalRes.finalText }],
145
195
  details: {
146
- runId: res.runId, todoId: res.todoId, agent: res.agent, model: res.model,
147
- status: res.status, durationMs: res.durationMs, tokenTotal: res.tokenTotal,
196
+ runId: finalRes.runId, todoId: finalRes.todoId, agent: finalRes.agent, model: finalRes.model,
197
+ status: finalRes.status, durationMs: finalRes.durationMs, tokenTotal: finalRes.tokenTotal,
198
+ retriedWithModel,
148
199
  },
149
200
  isError,
150
201
  };
151
202
  },
152
203
  };
153
- }
204
+ }