@agenit/cli 1.0.3 → 1.1.0

package/CHANGELOG.md

@@ -3,9 +3,310 @@
 All notable changes to `@agenit/cli` are documented here. The project
 follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-## [1.0.3] - 2026-05-07
+## [1.1.0] - 2026-05-08
+
+A minor bump because this batch is dominated by new user-facing
+flows: an interactive `/init` walkthrough, a Codex-style `/goal start`
+gate with plan generation + speckit pipeline, a Claude-style advisor
+mechanism the agent can call mid-tick, per-goal model pinning, and an
+end-of-session summarizer that finally populates the project memory
+files. None of these were in 1.0.5.
+
+### Added
+
+- **Interactive `/init` flow.** When `/init` runs without `--profile`
+  and the REPL is attached, a modal walks the user through:
+  - Profile picker (arrow-key nav over `automotive` / `embedded` /
+    `web` / `generic`, with the auto-detected option marked).
+  - `r` key triggers a Gemini classification call that scans the
+    project's high-signal files (`package.json`, `CMakeLists.txt`,
+    `Cargo.toml`, etc.) and recommends a profile.
+  - Custom-agent gate after profile selection — if Y, Gemini
+    drafts a project-tailored `SKILL.md` and writes it to
+    `.gemini/skills/<project>-agent/`. Verifier confirms Gemini-CLI
+    sees it on the next session start.
+  - GEMINI.md seeder added alongside AGENTS.md (Gemini-CLI's
+    auto-loaded persona file). Both files now include a
+    `## Recommended skills` section listing per-profile skills.
+- **Interactive `/goal start` gate.** When `[goal]
+  interactive_planning = true` (default), `/goal start` opens a
+  sequence of modals before spawning the runner:
+  - **Plan-first gate**: Y to plan via the speckit pipeline, N to
+    start the runner immediately, Esc to cancel.
+  - **Plan-model picker** (arrow-key nav): pick the model the
+    speckit specify+plan calls run on. Shows tier (smart / fast /
+    cheap / default) + the resolved model + provider per row.
+  - **Streaming plan loader**: speckit-specify runs first, then
+    speckit-plan. Phase headers stream in (`## phase: speckit-plan`)
+    so the user sees what's currently happening. Output is the
+    `plan.md` speckit wrote to `.specify/specs/<fid>/`.
+  - **Plan approval modal**: shows the on-disk plan.md. Y to
+    approve, E to drop into `$EDITOR` on the file, N/Esc to cancel.
+  - **Optional `clarify` gate** — runs the new two-pass clarify
+    (see below) when Y.
+  - **Optional `tasks` gate** — runs `speckit-tasks`, captures
+    `tasks.md`, injects it into every tick's prompt under
+    `## Task list`.
+  - **Worker-model picker**: same picker, different purpose —
+    pins the model the autonomous runner uses for every tick.
+  - The chosen models + speckit fid + tasks body live on
+    `GoalSpec` so the run is reproducible.
+- **Two-pass clarify flow.** The original `speckit-clarify` skill
+  expected interactive Q&A inside the model context, which doesn't
+  work in headless dispatch (it usually no-ops). The new flow:
+  - Pass 1: `extractClarifyQuestions` asks the model for ≤5
+    numbered questions about the spec. Streams.
+  - User walks the questions one at a time via a `ClarifyQAModal`
+    with text input. Esc skips a question; Ctrl-C cancels the whole
+    flow; empty Enter is treated as skip.
+  - Pass 2: `applyClarifications` sends the original spec + Q&A
+    pairs back, asks for a complete updated `spec.md` body, writes
+    it to disk. Streams.
+- **Advisor side-call mechanism.** The autonomous runner's main
+  model can emit `[[ADVISOR: <question>]]` mid-tick; the driver
+  detects the marker (top-level only — markers inside fenced code
+  blocks are ignored), side-calls a `tier: "smart"` GeminiClient
+  with the audit context, persists the reply to
+  `goal-turns/turn-NNN-advisor.md`, and injects it into the next
+  turn's user prompt under `## Advisor reply (consult #N)`. New
+  `[advisor]` config section: `enabled`, `tier` (default `smart`),
+  `max_consults_per_goal` (default 5), `redact_paths` (default
+  true — replaces absolute paths in the audit-note context with
+  `<path>` placeholders before sending to the smart model).
+  Recursive markers in the reply are stripped to keep the loop
+  terminating. Advisor cost entries land in `costs.json` with
+  `kind: "advisor"` (matching turn integer to the tick that emitted
+  the marker; the existing tick entries gain `kind: "tick"`). New
+  `/goal advisor [<turn>]` subcommand prints a stored reply.
+- **Per-goal model pinning.** `GoalSpec.workerModel` overrides
+  routing rules + `cfg.gemini.model` for every tick of the run.
+  Set via the worker-model picker; saved into `goal.json` so the
+  pin survives restarts.
+- **End-of-session summarizer.** New `[session]` config section
+  (`summarize_on_end`, `tier`, `min_turns`). On `/exit` / Ctrl-C /
+  `/quit` (or via the `/session summarize` slash command), the
+  REPL reads back the session's JSONL turns, asks the model for
+  durable signal in four buckets (`context` / `decisions` /
+  `requirements` / `soul updates`), and appends each non-empty
+  bucket to its target file: `memory/projects/<p>/{context,
+  decisions, requirements}.md` and the global
+  `<flowHome>/.flow/soul.md` (under a per-project section).
+  Resolves the long-standing "soul-keeper sees 0 bullets" symptom
+  — once `context.md` has content, soul-keeper's distillation
+  loop has work to do. Skipped for sessions shorter than
+  `min_turns` (default 3) so trivial lookups don't churn pro for
+  nothing.
+- **Streaming everything.** Plan generation, speckit phases
+  (specify / plan / clarify / tasks), advisor side-calls, and the
+  session summarizer all use `chatStreamJson` with text-delta
+  forwarding. The loader modals render the model's output live —
+  no more frozen spinner.
+- **Tier-aware routing primitive promoted.** `pickRoutedModel` and
+  the new `pickByTier` live in a shared `routing.ts` module
+  (extracted from `commands/goal.ts`). `GeminiClient` accepts an
+  optional `routingRules` config + per-call `tier` option;
+  `chosenModel(opts)` resolves model name without spawning. Four
+  call sites tagged with intent: `profile-recommend` (fast),
+  `App.tsx` free-form chat (fast), `plan-generation` (smart, when
+  used as a one-shot fallback), `agent-generate` (smart). The
+  remaining 8 hardcoded sites (squad helpers, debug, cve, etc.)
+  are flagged for a follow-up.
+- **Lazy memoised model probe.** `model-probe.ts` resolves the
+  fast / smart / cheap models per (account, flowHome) and caches
+  the result at `~/.config/agenit/model-cache.json` with a 7-day
+  TTL. New `--reprobe` flag busts the cache after a Gemini-CLI
+  upgrade or account switch.
+- **`/session` slash command.** `/session` shows status (id,
+  file, turns); `/session summarize` triggers the summarizer
+  manually mid-session.
+- **`/goal output [<turn>]` and `/goal advisor [<turn>]`.** Both
+  print stored per-turn artefacts (full assistant response or
+  advisor reply) so users can audit what the agent saw at any
+  step.
+- **Goal driver liveness updates.** During a streaming Gemini
+  call, the driver throttle-pushes "thinking …" snippets to the
+  rail card every ~1.5s. Card no longer goes stale (`?` glyph)
+  during long ticks; it shows actual text from the model.
+- **JobDetailPanel "Advisor consults" section.** Lists every
+  consult by turn number, with a hint pointing at `/goal advisor
+  <N>` for the body.
+- ~50 new tests across routing, model probe, advisor marker /
+  redaction / parser, prompt injection, session summary parse and
+  apply, clarify question parser, layout solver, and pool
+  snapshot fidelity. Total: 510 (up from 462 at 1.0.5).
+
+### Fixed
+
+- **Card "stale" override during ticking.** The right-rail card
+  was rendering the `?` (stale) glyph during legitimate long
+  Gemini calls because `now - lastEventAt > 60s` triggered even
+  while the model was actively producing tokens. Stale now only
+  applies to `active` / `blocked` states — `ticking` is exempt
+  and shows the spinner instead.
+- **Card width math.** Card content was overflowing because
+  `paddingX={1}` (2 cols) wasn't being subtracted from the
+  computed `innerWidth`. Fixed; "0s ago" no longer renders as "s
+  ago" with the leading character clipped.
+- **Plan-loader frozen spinner.** Plan generation used
+  `gemini.chat()` (blocking text mode) — the user stared at a
+  spinner for the entire 20-60s pro call. Now uses
+  `chatStreamJson` with deltas forwarded into the modal buffer.
+- **Speckit-clarify file detection.** The pipeline expected a
+  separate `clarification.md` file but speckit-clarify writes
+  back to `spec.md`. The detector now checks for a
+  `## Clarifications` section or moved mtime instead.
+
+### Changed
+
+- **`Session.end()` no longer just writes a meta marker.** When
+  `[session] summarize_on_end` is true and the session has enough
+  turns, exit prompts the user to run the summarizer first.
+- **`GoalSpec` schema additions** — `plan`, `workerModel`,
+  `speckitFid`, `tasksPath`, `tasksBody`, `advisorConsultCount`,
+  `lastAdvisorReply`. All optional; on-disk goal.json from 1.0.5
+  loads without changes.
+- **`CostEntry` gains optional `kind: "tick" | "advisor"`** so
+  `costs.json` can attribute advisor consults separately while
+  keeping the integer turn key.
+- **Plan persistence path** — when speckit runs, the plan lives
+  at `<project>/.specify/specs/<fid>/plan.md`. Approval modal
+  reads from there and the `[E]dit` action drops into `$EDITOR`
+  on that file directly (no sidecar copy). One-shot
+  `generatePlan` still writes to `.agenit-plans/` as a fallback
+  when speckit isn't used.
+
+### Out of scope (follow-ups)
+
+- True interactive Q&A passed through to a single `speckit-clarify`
+  skill call (the two-pass approach replaces that pattern).
+- Tagging the remaining 8 hardcoded model call sites
+  (squad helpers, debug, cve, stages, print-mode, speckit-internal,
+  etc.).
+- Auto-summarizer that walks every session JSONL in
+  `.flow/sessions/` for a one-time backfill on long-lived projects.
+- `/speckit chain <feature>` non-interactive variant for scripted
+  pipelines.
+
+## [1.0.5] - 2026-05-08
+
+### Added
+
+- **Background-jobs dashboard.** A new right-rail in the REPL shows
+  one card per active background worker (autonomous goal runner,
+  soul-keeper, future scanners). Each card surfaces the worker's
+  state (`active` / `ticking` / `blocked` / `done` / `aborted`),
+  progress (`12/50 turns`, `7 bullets`), last event, and age.
+  Width-adaptive: wide rail at ≥140 cols, narrow at 120-139, single
+  horizontal strip at 100-119, hidden below 80 (jobs reachable via
+  the new `/jobs` command).
+- **New `/jobs` slash command.** Lists / inspects / stops background
+  jobs. `/jobs` lists all, `/jobs <id>` shows detail, `/jobs stop <id>`
+  stops a pool worker. Goal runners route to `/goal stop` per the
+  existing API.
+- **Init verification.** `/init` now finishes with a verification
+  step that confirms (a) skills directory non-empty, (b)
+  `.gemini/settings.json` valid JSON with all hook paths resolvable,
+  (c) `AGENTS.md` present, (d) Gemini-CLI binary works, and (e)
+  `gemini skills list` reports the seeded skills including a
+  sentinel name (proves end-to-end discovery, not just file
+  presence). Output uses `✓ / ✗ / -` icons with `↳ fix:` hints on
+  failure. `--no-verify` skips the step.
+- **`JobSnapshot` + pool snapshot API.** `LoopWorker.snapshot()`,
+  `LoopWorkerHandle.getSnapshot()`, and `LoopWorkerPool.onChange()`
+  let any worker publish rich card data to the bus without coupling
+  to UI code. The TUI bus republishes pool spawns/evicts to the
+  rail automatically.
+- 31 new tests covering job-layout solver, TUI bus batching, pool
+  snapshots, goal-snapshot mapping, init-verify checks, and
+  hook-path rewrites.
 
 ### Fixed
+
+- **`/goal start` no longer spins on empty model output.** Four
+  intertwined bugs that caused goal runs to log dozens of zero-work
+  turns:
+  - The auditor ran `git status --porcelain` against
+    `cfg.flow.flowHome` (the install root) instead of the user's
+    project directory. It now uses `cfg.gemini.workingDir`, the
+    same directory the spawned Gemini-CLI subprocess writes to.
+  - When the user's project wasn't a git repo, the auditor
+    silently reported zero file changes. `/goal tick` and
+    `/goal start` now run `git init --quiet` if `.git/` is
+    missing.
+  - The Gemini stream-json parser silently dropped responses when
+    the CLI used the newer `{type:"content", text:"..."}` event
+    shape instead of the older `{type:"message", role:"assistant",
+    delta:true, content:"..."}`. Both shapes are now accepted.
+    Empty responses on clean exit (auth errors, schema drift) now
+    throw an error including stderr + a stdout sample, instead of
+    looping indefinitely.
+  - `/goal status` reported `runner: running` after the autonomous
+    loop had exited on its own (completion / budget exhaustion /
+    error). The `activeRunner` reference now nulls itself when the
+    `stopped` event fires.
+- **Stale relative hook paths in existing `settings.json` are now
+  refreshed.** A user who ran `/init` before the 1.0.4 path-rewriter
+  shipped had `./.flow/hooks/foo.py` tokens that fail at hook-fire
+  time because the user's project doesn't ship `.flow/hooks/`.
+  Re-running `/init` now re-rewrites those tokens to absolute paths
+  under flowHome while preserving any user customisations elsewhere
+  in the file. The fix is idempotent and only writes when something
+  actually changed.
+- Backend errors during a goal tick (auth failures, model rejections)
+  now write an audit row with the error in `note` + `errors[]`,
+  mark the goal `Aborted`, and stop the runner cleanly. Previously
+  the runner could die without trace.
+
+### Changed
+
+- `LoopWorkerHandle.done()` and `.stop()` now resolve only after
+  pool eviction completes — `pool.list()` is consistent with
+  `await handle.done()`.
+- `tuiBus` updates are coalesced via microtask flush so producers
+  firing many concurrent `setJob()` calls cause one re-render per
+  event-loop tick instead of strobing the rail. `flushSync()` is
+  exposed for deterministic tests.
+- `tuiBus.setJob()` and `removeJob()` now allocate a fresh
+  `backgroundJobs` array on every change so React's referential
+  equality fires for memoised consumers.
+
+## [1.0.4] - 2026-05-08
+
+### Added
+
+- **`agenit init` now seeds `.gemini/` for the user's project.** Copies
+  every skill from `<flowHome>/.gemini/skills/` into
+  `<project>/.gemini/skills/` and writes a `<project>/.gemini/settings.json`
+  derived from TheFlow's, with all relative hook paths
+  (`./.flow/hooks/foo.py`) rewritten to absolute paths so the hooks
+  fire correctly regardless of cwd. Without this step the cwd fix
+  introduced in 1.0.3 left Gemini-CLI unable to discover any skills or
+  hooks when running from the user's project — `/feature` runs would
+  silently degrade to a generic persona. Use `--no-skills` to opt out;
+  an existing `<project>/.gemini/settings.json` is preserved.
+- New `tests/gemini-assets.test.ts` covers skill copying, hook-path
+  rewriting, and existing-settings preservation.
+
+## [1.0.3] - 2026-05-08
+
+### Added
+
+- **Cross-platform Gemini CLI discovery.** `binary` in `flow.toml` can
+  now be a bare command name (`"gemini"` is the new default) — agenIT
+  searches `PATH` (with `PATHEXT` extensions on Windows) and falls
+  back to well-known install dirs per OS:
+  - macOS: `/opt/homebrew/bin`, `/usr/local/bin`, `~/.npm-global/bin`
+  - Linux: `/usr/local/bin`, `/usr/bin`, `/home/linuxbrew/.linuxbrew/bin`,
+    `~/.npm-global/bin`, `~/.local/bin`
+  - Windows: `%APPDATA%\npm`, `%ProgramFiles%\nodejs`,
+    `%LOCALAPPDATA%\Programs\gemini-cli`
+
+  Absolute paths are still honoured verbatim. On Windows the spawned
+  process uses `shell: true` so `.cmd` / `.bat` shims execute correctly
+  when invoked without their extension.
+
+### Fixed
+
 - **Artefacts now land in the user's project, not in the install repo.**
   Previously every relative path in `flow.toml` (gemini `working_dir`,
   `memory_dir`, `codedigest_dir`, `mempalace`, sona/reasoning/rvf)
@@ -25,29 +326,18 @@ follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   shows "Spec-Kit (specify → plan → tasks)" and "Export spec → SWE.1"
   before the V-Model nodes start firing their own stage updates.
 
-### Added
-- **Cross-platform Gemini CLI discovery.** `binary` in `flow.toml` can
-  now be a bare command name (`"gemini"` is the new default) — agenIT
-  searches `PATH` (with `PATHEXT` extensions on Windows) and falls
-  back to well-known install dirs per OS:
-  - macOS: `/opt/homebrew/bin`, `/usr/local/bin`, `~/.npm-global/bin`
-  - Linux: `/usr/local/bin`, `/usr/bin`, `/home/linuxbrew/.linuxbrew/bin`,
-    `~/.npm-global/bin`, `~/.local/bin`
-  - Windows: `%APPDATA%\npm`, `%ProgramFiles%\nodejs`,
-    `%LOCALAPPDATA%\Programs\gemini-cli`
-
-  Absolute paths are still honoured verbatim. On Windows the spawned
-  process uses `shell: true` so `.cmd` / `.bat` shims execute correctly
-  when invoked without their extension.
-
 ### Changed
+
 - Default `flow.toml` ships `binary = "gemini"` (was the macOS
   Homebrew absolute path) and drops the leading `../` from
   `memory_dir` / `codedigest_dir` so artefacts land *inside* the
   user's project.
 
 ### Tests
+
 - New `tests/config-paths.test.ts` pins the install vs project anchor
   split and the absolute-path passthrough.
 - New `tests/binary-resolver.test.ts` covers PATH walking, absolute
   path probing, and missing-binary handling.
+- New `tests/gemini-assets.test.ts` covers skill copying, hook-path
+  rewriting, and existing-settings preservation.

package/README.md

@@ -39,9 +39,9 @@ agenit run
 ## Commands
 
 | Command | Description |
-|---|---|
+| --- | --- |
 | `agenit` | Interactive REPL |
-| `agenit init [dir]` | Scaffold project memory and `.agenit_project` marker |
+| `agenit init [dir]` | Scaffold project memory + `.gemini/` assets + `.agenit_project` marker. Verifies Gemini-CLI can see the seeded skills before returning. Pass `--no-verify` to skip. |
 | `agenit run` | Run the full V-Model pipeline |
 | `agenit audit` | Show traceability registry |
 | `agenit projects` | List all projects in memory |
@@ -49,6 +49,14 @@ agenit run
 | `agenit sessions` | List recent sessions |
 | `agenit completions <shell>` | Print shell completion script (bash / zsh / fish) |
 
+Selected REPL slash commands (full list in `/help`):
+
+| Command | Description |
+| --- | --- |
+| `/goal <objective>` | Codex-style autonomous loop. `/goal start` runs ticks back-to-back; the right-rail card shows turn budget + last decision live. |
+| `/jobs` | List active background jobs (goal runner, soul-keeper, scanners). Useful when the terminal is too narrow for the right-rail. |
+| `/init` | Re-seed `.gemini/` skills + settings; verifies Gemini-CLI can see the result. |
+
 ## Configuration
 
 Place a `flow.toml` (or `agenit.toml`) in your project root, or pass `--config <path>`. See the [project documentation](https://github.com/muhammed-eldabea/flow) for the full schema.