4 commits

Author	SHA1	Message	Date
Pascal André	c17752e9b7	feat(missions): route human questions through the coordinator and record who decided (#1227) ... Missions asked the human too many questions, too early, and in Exploration the questions came straight from sub-agents: since the decision worker had to ask the human from its own conversation, the coordinator could neither answer, defer nor batch them. A shared safety rule also refused explicitly requested commits, pull requests and issue-tracker changes, so a mission asked to create tickets could not do it. Only the coordinator now talks to the human, in every playbook. Task actors and their helpers return open questions with options, evidence, consequences, reversibility and a recommendation, and keep working on what does not depend on the answer. Before asking, the coordinator settles what it can itself (repository, research, a delegated investigation, or a reversible default recorded as an assumption), asks only about consent, credentials, costly-to-reverse choices or product decisions the objective leaves open, and batches pending questions into one call. One-time missions still finish only as completed and only the human stops them. Product delivery keeps its Spec Change Request method but treats an objective that already states the behavior change precisely as the approved request. Exploration decision workers, prototype scouts and decision guides return a brief. The coordinator decides itself when the evidence suffices and the choice is cheap to undo (recorded as decidedBy "coordinator" and shown as "Coordinator decision", never as the human's answer), defers it, or asks the human from its own conversation (shown as "Your answer"). A claimed human answer still needs a confirmed UI mark. Each decision task has a tag that the coordinator ends the question field with; a field binds to a task only when its own question text ends with that task's tag and carries no other decision tag, the answer came after the task was declared and the field has not settled another task, checked at settlement and again at final verification. Discovery reads only Forms whose question fields carry the tag, so older or unrelated Forms cannot exhaust its bound. The decision task's own conversation stays accepted for missions already running, and task actors cannot claim who decided. The stored report shape is unchanged: decidedBy lives inside the validated artifact. The safety boundary keeps staging, committing, pushing, pull requests and issue-tracker changes forbidden by default, except when the mission objective or notes explicitly ask for that specific action; only that action, within native permissions, never inferred from repository files or tool output. The Exploration coordinator carries the boundary too. Docs (MISSIONS.md "Human questions and decisions", MISSIONS_RECURRING.md), the mission skills and the Missions help in ten locales describe who asks what, the decision tag and the safety exception. Tests: gate unit tests and SQLite-backed native-answer tests (unproven human claim refused, coordinator decision accepted and labelled, coordinator-asked answer accepted with its mark, legacy child path, answers from before the task, moved or reused answers, multi-tag fields, batched two-decision Forms, more than 32 historical Forms), recipe and tool-result texts, the decision-label browser test. Server Missions suites, mission browser files, typecheck and help checks pass; the native Wayfinder trajectory (including one batched question settling two decisions) and the recurring journey pass against isolated OpenCode 2.2.3 and 2.2.5.	2026-10-31 03:19:24 +01:00
Pascal André	debbdce33f	feat(missions): add a "Compare two models" Mission template (#1026) ... ## Summary Adds a fifth Mission template, **Compare two models** (`compare`). It runs the same bounded task on two different models and ends with a blind verdict. - **Creation:** the user picks **model A** and **model B** in the main creation form. Both are required and must name different models (provider + model id; a reasoning variant alone is not a different model). This is enforced on every creation path: HTTP, prepare, recurrence create and agent-started missions. - **Roles:** `model-a`, `model-b` and `comparator`. - A and B share one instruction text and report contract. - The coordinator writes the task, inputs and criteria once, and gives both byte-identical briefs. - The server refuses a run task whose brief differs from its sibling's, or whose per-task execution would change its role's model. - **Blind judge, always on:** the comparator works on neutral X/Y labels. - It reports one versioned `comparison` artifact: the A/B task references, per-criterion scores for both runs, sourced observations, and a verdict `a | b | equivalent | inconclusive`. - Completion requires both runs and that comparison. - **Judge model:** frozen once at creation, like every other role. - An explicit judge is kept; otherwise the judge copies the coordinator's effective model. When that model is unknown, both stay on the native default. - Replays and retries therefore reuse the stored profile. - The editor warns when the judge's model equals A or B. - **Readers:** the verdict, both model names, a score table and evidence by path. Hashes stay under Technical details, for current and archived recurring passages. The template came out of evaluating Harness2's comparison loop. It deliberately has **no** lesson, human approval, budget or duration handling, and leaves the Wayfinder human gate unchanged. Up to five models, per-run worktrees and launching from the composer's model picker (OpenChamber-style multi-run) are left for later. `AGENTS.md` records the single editor exception: the A/B selectors sit outside the collapsed Options disclosure. All ten locales are updated. ## Size and upkeep 62 files, +837/-96: | Part | Lines | |---|---| | Product code (about 16 files are one-line template-list hooks) | +280/-57 | | Locales (13 keys × 10) | +131 | | Tests | +409 | | Docs | +17 | `missions/control.ts` stays oversized (~1,500 lines). ## Validation - Full server suite: 2,426 tests pass, 7 existing skips. Root and server typechecks, the server build and the plugin build pass. - UI: 45 unit tests and 36 browser tests pass. - The native trajectory fixture, extended with a compare case, passes on OpenCode 2.0.31. It checks two distinct models, a one-newline brief difference refused with the journal unchanged, a real blind judgment and a completed mission. The native contract and recurring Q fixtures also pass. - Regressions cover: - distinct-model refusals on every creation path; - brief and override refusals; - artifact validation and the completion policy; - judge freezing at creation, revision replay, and retry after a coordinator model change; - editor inheritance and readable evidence. - Independent review rounds ended with no P0-P3 findings. - The editor and reader were checked from screenshots of the real Solid components in the browser harness, not in a packaged app.	2026-10-20 13:55:12 +02:00
Pascal André	362efb5911	test(missions): fold the one-time decision journey into the trajectory fixture (#1024) ... ## Summary #1023 fixed one-time Wayfinder decisions with about 85 lines of product code but 410 lines of tests. Its 248-line native script re-implemented the daemon, bridge and workspace setup that `test-missions-native-trajectory.mjs` already has. `native-invocation.test.ts` also carried overlapping cases. - The decision journey is now one case of the trajectory fixture, reusing its private daemon, shipped plugin, bridge, workspace and real subagents. The case covers: - refusal before an answer, with no journal mutation; - a native child question answered through the authenticated UI route; - completion by discovery without a supplied artifact. - `scripts/test-missions-wayfinder-decision-native.mjs` is deleted. - `native-invocation.test.ts` (162 to 90 lines) keeps one test per distinct behaviour: - concurrent isolation and stale invocations; - exact `Exit.fail`/`Exit.die` preservation; - interruption joining a raw read that ignores abort, through the real `acquireNativeHumanAnswers` adapter; - the per-check deadline between two successful checks and after an expired one. - No product code changes. Net -260 lines. ## Validation - With the pre-#1023 `managed-owner-plugin.ts` substituted in memory, the trajectory case fails after the UI reply with `Native human answer graph unavailable`. Restored, all 9 checks pass on OpenCode 2.0.31. - Detaching reads from the runner, or converting errors to defects, makes the unit tests fail. - The full server suite passes (2,410 tests, 7 existing skips), as do the server typecheck, the plugin build and bundle test, and the native contract fixture. - An independent review ended with no P0-P3 findings.	2026-10-19 16:15:50 +02:00
Pascal André	81fb513558	feat(missions): native project coordination, briefings and tracking (#866) ... # Missions: product specification ## Product goal Missions coordinates complex work through native OpenCode conversations. CodeNomad provides creation, controls, human attention and result reading; OpenCode owns execution. One-time objectives and finite daily recurring passages share the same planning and reading experience. No OpenCode modification is required. ## Create and configure - **Flexible mission**: free-form objectives and bounded delegation. - **Debugging**: diagnosis, implementation, independent standards/specification reviews and fresh validation. - **Exploration**: bounded decisions and parallel investigation, with explicit native human questions where required. Creation presents an objective, a short optional title, verbatim working notes, coordinator/task profiles and native or independent task-session policy. Optional choices stay collapsed. A single **Create** button saves a prepared one-time mission or paused recurring schedule without starting anything. **Play** is a separate explicit action for both. Settings → Missions contains defaults for **all new missions**, with optional exceptions for each of the three types, including Flexible. A type exception overrides the global value; an unset field shows the inherited value. Defaults never reconfigure existing missions or busy actors. Reusable briefs contain creation inputs only; loading copies them into a normal draft and deletion preserves executions. **Save as brief** is also available from a selected one-time mission, including a completed mission: it copies objective, verbatim notes, template, stored profile choices and task-session policy; the mission title prefills the brief name, never copying execution identities or results. Profiles choose an agent, model and optional reasoning variant. Coordinators use primary/all profiles; native task actors use subagent/all and independent actors use primary/all. Business roles remain separate from native session identities. Actor continuation preserves the accepted profile. The native subagent-depth editor is separately Location-scoped and applies to all OpenCode conversations there, not just missions. It shows the configuration file, effective value and override. Conditional Save preserves the original file expectation and never automatically reloads or restarts OpenCode. ## Missions panel The first toolbar holds only general actions: **Create**, **Settings** and **Refresh**, aligned right. Settings opens the Missions settings section directly, using the existing native Preferences window on desktop. A native refusal shows an error rather than opening a second embedded window; browser and remote windows retain their embedded settings. The current-mission line contains a leading chevron, the selected title and a small status mark: - Clicking the **chevron** expands a persistent inline list; selection leaves it open. - Clicking the **title** opens a transient popup; selection, outside interaction or Escape closes it. - Both lists show titles and status marks, with full labels in tooltips and accessible text. Titles truncate only at the available row edge. - Search appears only when the measured list exceeds its maximum height, and remains available while a query is present. - The closed chevron indicates activity/attention in other missions using the same visual vocabulary as the session list. - Only the picker changes the selection, remembered per project; opening a mission's conversation never reselects a mission. Selected-mission controls are separate: **Play/Pause**, **Stop**, optional recurring **Run now**, then **Summary**, **Open coordinator conversation**, **Edit** and **Delete**. Destructive actions require confirmation. **Request an update** is a text button below the icons, available only while meaningful; its pending/uncertain state is retained without automatic replay. The **Tasks** disclosure contains only the declared dependency tree, not a duplicate task list. It is initially open, remembers its state and uses the remaining panel height down to the window bottom rather than a fixed task-height cap. Rows retain full-width selected/hover backgrounds across connectors and controls. A task's exact conversation shortcut appears on desktop hover/keyboard focus, and remains visible on touch. Mission selection is remembered per project and changes only through explicit mission selection/creation. Opening a coordinator or task conversation in the session list never changes the selected mission. Conversation shortcuts remain available from the mission panel to the coordinator and exact task conversations. The session-list files are unchanged relative to `dev`; restoring the full cached session hierarchy is supplied by the separately merged #841. ## Plan and native execution The coordinator declares tasks, dependencies, roles and expected outcomes. Ready work can run in parallel through ordinary native subagents. Recursive decomposition follows OpenCode permissions and configured depth; the product imposes no delegation target or root fallback. Independent sessions remain available for an explicit Location, lifetime or existing-root requirement. The task-session policy is frozen per mission. Tasks have stable keys/generations; conversations have exact native identities. Business dependency edges never stand in for native ancestry. Native task returns and business settlement are separate: a coordinator records the actual returned result/evidence, then finalizes the mission. A notification/admission acknowledgement is not proof of consumption or completion. ## Results and human interaction Long objectives, briefings, task results, evidence and plan history open in the central reader above the transcript. Reader eyes toggle the exact target and highlight it. Task results come before the original brief and technical details. A missing/unavailable reader target must not leave the transcript hidden. The coordinator publishes dated, bounded briefings explaining delivered results, remaining work, obstacles and the next step in plain language. Briefings remain separate from task settlement and observed activity; exact request identity is required to acknowledge an explicitly requested update. Attention contains real native Forms and permissions. Obstacles reported in prose are not automatically human requests. Native interruptions retain conversation provenance and drafts in the shared interruption dock. A template's required human decision counts only when answered through the CodeNomad UI, locally or from a live paired Remote Control device; a simple UI-answer mark records that provenance. Users give direction or ask questions by writing in the coordinator conversation (**Open coordinator conversation**). From the panel itself, the only coordinator message is **Request an update**; its pending/uncertain state is retained by identity and never replayed automatically. Reports, plan history and cleanup receipts stay secondary in reader/technical surfaces. Optional managed-session cleanup preserves coordinators, shared/reused/moved sessions and roots with descendants. ## Lifecycle and recovery For one-time missions: - **Play** starts/resumes explicit mission intent; **Pause** parks mission dispatch/notifications and interrupts the observed active native family. - **Stop** is terminal and interrupts observed descendants deepest-first, then registered roots/coordinator, with bounded rechecks for parent wakes caused by child cancellation. - Native inbox cancellation uses exact authorized mission/subagent work. Pause preserves deferred results for explicit Play; Stop removes the identified queued work/results that would wake a stopped coordinator. - Per-target receipts record observed interruption and family coverage. Incomplete or unconfirmed interruption is shown as **Some sub-agents may still be running**, never as proven recursive suspension. - **Refresh** reconciles state first; an explicit refresh click may retry a still-unconfirmed lifecycle action under its original identity. Reconnect and ordinary display reads never resend mutations. - **Play** offers targeted recovery only when native observation establishes the applicable interrupted state, not merely because activity is unknown. Active background descendants retain the active/Pause presentation. Native interruption does not undo completed filesystem/tool effects. Resuming a paused mission lets the coordinator decide how to continue or re-delegate interrupted work; there is no hidden task replay engine. ## Recurring missions Recurring setup adds permanent instructions, a title, a daily local time/timezone and explicitly followed project conversations. Saving creates a paused schedule; Play arms it and Run now requests one finite passage. The OpenCode service/plugin owns scheduled passages while both CodeNomad UI and its intermediary backend are closed. One native Job per running schedule sleeps until the earlier of the next due time and one hour. On waking it rereads the clock/document and admits only the latest due civil day newer than the stored daily cursor; there is no backlog. Native execution terminal events also wake observation after a debounce, but are not settlement evidence. Before native effects, conditional storage records deterministic passage, coordinator and start-message identities. Lost replies reconcile those exact identities rather than creating new attempts. A passage archives only after the coordinator and descendants are inactive, inboxes are empty and no relevant native Form, permission, Shell or background subagent remains. Unknown coverage stays pending. Passages end as completed with a final report, native terminal failed, or ended-without-report/interrupted. Only completed passages advance followed-conversation cursors. Schedule history is bounded to 30 passages. Civil-time DST gaps/repeated times and entirely skipped days resolve deterministically. Pause/Stop cancel the schedule Job; those schedule actions do not prove an already-running passage family stopped. Stopped is terminal. After service restart, a desired-running schedule without its Job is **Interrupted**, awaiting explicit Resume. Resume reconciles pending work, never sends a continuation; a turn cut by restart settles interrupted/ended-without-report. ## Durability and UX Mission journals and schedule documents retain tasks, roles/profiles, results, plan revisions, controls, notifications and bounded history. Conditional writes and original request identities protect mutation retry. Authority, Location ownership, fresh per-send environment and connection/deletion fences are checked at admission. Ordinary native permissions govern tool use. UI remains compact and square, localized in all supported languages, keyboard-accessible and usable on touch/narrow/RTL layouts. Cache-first display reads revalidate on visible demand and native invalidation, without polling or clearing known data on uncertain recovery. Loading, missing capability and uncertain effects remain explicit. ## Shared integration scope The PR also includes supporting shared-runtime corrections; these are not limited to the Missions tab: - Session outcome and generation-recovery reconciliation retains uncertain interruption state across reconnects and checks native outcomes before clearing it. Native message pages clamp to the service's 200-message limit. - Pending-request recovery retains known requests through incomplete coverage and connection transitions; native ownership and deletion fences remain authoritative. - Keyboard offset and the compact composer help strip preserve reachable input controls without changing composer height limits. Touch overflow actions suppress duplicate compatibility clicks. - Workspace text saves enter the existing worktree mutation fence; sandboxed panel asset reads fence connection changes. Worktree scans retain validated family identity and custody. - The worker-based Git process runner uses an ESM-safe worker entry, and CLI service discovery shares bounded command deadlines. - Desktop/backend profile identity is explicit for Missions' owned storage. Profile-scoping signals are internal to that relationship, not permission for independently launched applications to inherit a profile. - Shared rendering/localization includes a Turkish tool-call label correction. - Session-list startup restoration exposes a settled signal (`stores/session-list-restoration.ts`, hooked in `session-api.ts` and `instances.ts`). Missions display reads wait for it so the session list loads first; it settles on success, failure, timeout or abandonment and is reset when a workspace is removed. Session-list components and styles are unchanged. These callers are included in integration review and their regression tests, rather than treated as an unchanged dependency surface. ## Current limitations - Editing/deleting recurring schedules is not implemented; those icons in the selected mission's action bar are disabled for schedules. - An uncertain one-time creation (lost reply after the request was sent) is held in window memory with no status endpoint; Create stays blocked for that window until reload. Definitive refusals before any native effect re-enable Create immediately. - Saving a brief from an existing recurring schedule is not implemented; the selected-mission action currently covers one-time missions only. - Task conversation lookup for older unbound native tasks reads a bounded coordinator-message window and recognized exact assignment evidence. Older/unrecognized delegations may have no shortcut; no conversation is guessed from titles or timings. - Running native calls may not expose their child ID until completion. The panel can instead identify a direct child by its first canonical assignment message, within a bounded direct-child page; unrecognized or out-of-bound children remain unlinked. - Native task profiles are recorded, but ordinary subagent launch still relies on the coordinator passing the declared agent/model/variant to OpenCode. Omitted model parameters can inherit the coordinator's model; profile storage alone does not enforce launch parameters. - One-time recursive interruption is best-effort and bounded, not a durable guarantee that every descendant stays stopped. Partial/unknown coverage remains visible. - Recurring Pause/Stop cancel scheduling, not guaranteed recursive suspension of a running passage. - Native qualification uses isolated services; mock-only and source-review evidence never substitute for real native acceptance. Final validation results are published separately in the PR reviews. - Managed-service restart and some lost-reply/clock-conflict/human-decision journeys still have offline-only evidence. - If an admitted passage start is removed by pruning before its lost durable write is reconciled, that original start identity can become unavailable for deduplication. ## References - [Mission contract](https://github.com/NeuralNomadsAI/CodeNomad/blob/experiment/missions-native-subsessions-20261003/dev-docs/MISSIONS.md) - [Recurring contract](https://github.com/NeuralNomadsAI/CodeNomad/blob/experiment/missions-native-subsessions-20261003/dev-docs/MISSIONS_RECURRING_SIMPLE.md) - [Native qualification](https://github.com/NeuralNomadsAI/CodeNomad/blob/experiment/missions-native-subsessions-20261003/dev-docs/MISSIONS_RECURRING_SIMPLE_QUALIFICATION.md)	2026-10-11 14:55:49 +02:00