Skip to content

V0_8_61_TRAIN_PROMPTS.md

Latest commit

 

History

History
139 lines (122 loc) · 9.62 KB

V0_8_61_TRAIN_PROMPTS.md

File metadata and controls

139 lines (122 loc) · 9.62 KB
Raw

v0.8.61 — six parallel worktree prompts (one per train)

The runtime-control-plane work splits into 6 independent trains. Run each in its own git worktree (or hand each prompt to one agent with isolation: worktree). Trains touch mostly disjoint subsystems, so they parallelize; within a train, work the issues in the given order. Land each train as its own branch/PR against codex/v0.8.61.

Thesis (don't reorder): runtime control-plane correctness leads — Trains 1→2→3→4. Train 5 (UI) and 6 (distribution) ride along and can run anytime. Train 4 (goal mode) depends on Train 3 (durable nonblocking workers) being real.

Common preamble — PREPEND this to each train prompt below

You are implementing one workstream of CodeWhale v0.8.61 in an ISOLATED git worktree.
CodeWhale is a Rust workspace (provider-agnostic terminal coding agent; first-class
DeepSeek + GLM/MiniMax/Moonshot/OpenAI/etc).

WORKTREE CONFINEMENT (do FIRST — this prevents a real bug we hit):
1. Run `pwd` and `git rev-parse --show-toplevel`; confirm you are inside a worktree
   (path contains `.claude/worktrees/`). That toplevel is YOUR repo.
2. Edit files ONLY under that toplevel (relative paths). NEVER write to an absolute
   path outside it (e.g. /Volumes/VIXinSSD/codewhale/crates/...) — that is a SEPARATE
   checkout and writing there corrupts it.
3. Before committing, `git status` and confirm every changed path is inside your worktree.

CONTEXT — read these first (they are in your worktree if it branched from codex/v0.8.61;
otherwise read the issue + the design doc the orchestrator pasted):
- docs/V0_8_61_ISSUE_COVERAGE.md — disposition + plan for every milestone issue.
- docs/V0_8_61_DESIGNS_BATCH1.md — code-ready designs for several of these issues.
- docs/AGENT_RUNTIME.md, docs/SUBAGENTS.md — the one-runtime model.

FOUNDATIONS ALREADY LANDED (build ON these, don't duplicate; several are #![allow(dead_code)]
awaiting their first consumer):
- worker_profile.rs: WorkerRuntimeProfile (role/permissions/shell/tools/model-route/depth/
  background) + derive_child() non-escalating intersection. (#3217/#3211/#3213/#414/#426/#1186)
- goal_loop.rs: decide_continuation() — wired into turn_loop goal continuation. (#3215)
- crates/state record_thread_goal_usage(): durable per-goal token/time accrual.
- model_registry.rs (#3071/#3073), provider_readiness.rs (#3083), context_budget.rs (#3086),
  provider_adapter.rs (#3084), resource_telemetry.rs (#2666), request_tuning.rs (#3024).
- Freeze fixes already landed: cancel-between-batches; parent no longer barriers on running
  sub-agents (should_hold_turn_for_subagents). Route isolation #3227 landed.

ETHOS (AGENTS.md): preserve first-class DeepSeek support + CodeWhale branding; treat
community work as evidence; review from code+tests not titles; positive crediting comments
only; do NOT tag/publish/release or merge/close without the maintainer's explicit approval.

RULES: keep each change scoped + safe; do NOT weaken/remove existing tests or touch
Cargo.toml/lock/version/README/CHANGELOG version files. Everything must COMPILE and your new
tests must pass: `cargo fmt`, then the focused tests for what you touched. Do NOT run
`cargo build --release` (CI does). Commit incrementally with messages ending `Refs: #N`.
If an issue is epic-scale or unsafe in one pass, land the smallest safe slice + tests and
mark the rest as a clearly-noted follow-up rather than forcing a broken change.

OUTPUT: per issue — status (implemented/partial/blocked), files, the exact test command +
result, a short summary, and risks. Then the branch name + head SHA(s) for harvesting.

Train 1 — Route / model isolation (LEAD — bad routes poison every test)

Issues: #3227 (DONE), #3205, #3204, #3213, #3071 (registry seeded), #3072, #3073, #3075, #3024 (support-map seeded), #3025, #2027, #1768. Outcome: one route-effective model service; session-local provider/model state; correct context-window metadata + preflight over-limit; per-role/scout-vs-synthesis model routing. Guidance: build the route-effective model inventory/service (#3205) on model_registry

  • provider_readiness + provider_adapter; fix context-window metadata + over-limit preflight (#3204) using context_budget; split model-facing capabilities from human mode labels (#3213); migrate hard-coded model lists to model_registry (#3073) and hydrate it with an offline cache (#3072); wire per-role + scout/synthesis model routes (#2027/#1768) through worker_profile::ModelRoute and request_tuning. Add a route-resolution test matrix.

Train 2 — Permissions + shell nonblocking

Issues: #3211, #3212, #1186, #2475, #1791, #1786, #1737. (+ harden the pre-existing env-mutating config tests, e.g. save_api_key_for_openrouter_writes_provider_table does unsafe set_var without serialization — add a serial guard / env restore.) Outcome: first-class permission profiles; foreground/background shell semantics; read-only shell parallelism; no pointless wait-blocks; typed persistent permission rules. Guidance: wire worker_profile (PermissionSet/ShellPolicy) into the real approval/shell path, replacing the shell boolean (#3211/#3217-permissions/#1186); default independent shell + verifier work to background jobs (#3212); make synchronous tools (file_search/ grep_files/list_dir) cancellable so they don't block turn cancellation (#1791); reconcile failed-shell stuck state + PID/work-queue hangs (#1737/#1786); fix the YOLO/MCP prompt interruption (#2475). See design batch for #3212/#1786/#1737.

Train 3 — Worker / fleet / sub-agent convergence (the heart of 0.8.61)

Issues: #3096, #3154, #3166, #3167, #3216 (Bug A done), #3217 (profile seeded), #3226, #2211, #1806, #1679, #2652, #719, #414/#426 (intersection seeded). Outcome: "sub-agent" becomes UX language over fleet-style durable workers; the parent stays responsive; workers have rich states, receipts, retries, tool profiles, and a parent-visible interaction contract. Guidance (the two highest-value fixes, from the convergent analysis):

  • Bug 1 (freeze): decouple the input loop from the engine-event drain — move the blocking event::poll()/event::read() (ui.rs ~2859) onto its own thread feeding a channel; AND coalesce AgentProgress events to one redraw per agent per drain.
  • Bug 3 (the hang): adopt the autonomous-to-completion contract — a sub-agent must never park on input_rx.recv() (WaitingForUser). If it lacks info it makes a sensible default and records the assumption, OR terminates with a structured NeedsInput { question } result on the existing completion channel, so the parent wakes and re-dispatches. Keep agent_eval send as best-effort steering only.
  • Bug 2 (status): plumb the real AgentWorkerStatus (10 states) to the sidebar instead of hardcoding "running" (sidebar.rs ~2077); widen the tool-facing SubAgentStatus so agent_eval/agent_open expose Queued/ModelWait/WaitingForUser. Add #3226's parent-visible worker interaction contract (recommended action per worker).
  • Wire agent_open to build a WorkerRuntimeProfile + (durable mode) enqueue a fleet worker run per docs/AGENT_RUNTIME.md (#3096/#3154/#3217). Fleet dogfood smoke (#3166), org chart (#3167). Add the six-worker stress harness asserting input/render/cancel stay live (#3216).

Train 4 — Goal mode (before /swarm) (depends on Train 3 durable workers)

Issues: #3215, #3218, #891, #1976, #2058, #2029. Outcome: a persistent objective loop that survives turn boundaries, resumable + durable, with visible token/time accounting; /swarm stays gated until the substrate is real. Guidance: make goal_loop cross-turn — lift the continuation counter to durable state on the ThreadGoal store, increment usage via record_thread_goal_usage, and re-dispatch a worker turn toward the objective until decide_continuation says Stop. Bridge the three goal models (HuntState / runtime GoalState / durable ThreadGoal) onto the durable one. Add a verifier-as-judge gate before update_goal complete (#2058). Keep /swarm gated (#3218) until Train 3 lands. Sub-agent checkpoint/continue across turns (#2029).

Train 5 — Composer / steering / TUI clarity (parallel; rides along)

Issues: #3203, #3224, #2054, #3194, #3188 (DONE), #3190, #2982, #3028, #3078, #963, #2666 (telemetry seeded). Outcome: reliable queued steering + Ctrl+S; discoverable/configurable shortcuts; clear busy/free state; clickable stop/inspect affordances; word-wrap; live token throughput. Guidance: reliable queued steering + Ctrl+S send + Ctrl+Enter terminal-conflict handling (#3203/#3224); composer queued-steer state labeling with row send-now/drop/clear (#2054); helper-hint audit (#3194); a visible busy/idle indicator in the footer (#2982); click-to-act on sidebar rows + stop targets (#3028); auto-clear completed sub-agent cards with TTL (#3078); word-wrap truncation fix (#963); surface token throughput during streaming using resource_telemetry (#3190/#2666).

Train 6 — Distribution + release hygiene (separate hardening train)

Issues: #3207, #3208 (DONE), #2960, #2924, #2917, #1067, #3214 (DONE), #3192. Outcome: clean install/update paths; correct Linux artifact naming + glibc; branch hygiene; ACP registry submission. Guidance: Linux glibc requirement + GLIBC_2.39 not found (#3207/#1067) — build against an older glibc or document the floor; rebrand update path (deepseek update/npm) (#2960/ #2924/#2917) — extend the legacy-binary detection already in crates/cli/src/update.rs; ACP registry submission to agentclientprotocol/registry (#3192) — primarily an external PR + adapter-completeness review. branch-hygiene.sh (#3214) already shipped.