atom-agent 1.2.0 → 1.3.0

package/CHANGELOG.md

@@ -1,5 +1,80 @@
 # Changelog
 
+## 1.3.0 — 2026-09-11
+
+- Session goals (`src/goal.ts`, `src/App.tsx`, `src/agent/loop.ts`,
+  `src/agent/goal-evaluator.ts`): `/goal <objective>` pins one session
+  objective that runs turn-to-turn with no turn cap until paused, cleared,
+  or a `complete`/`blocked` verdict. Bare `/goal` shows text, state, and
+  cumulative stats (turns · requests · tokens · work); `pause` halts with
+  everything kept; `resume` re-arms (idle starts a continuation turn, busy
+  resumes at turn end); `clear` ends it. Cancel and spent step/tool-call
+  budgets pause, never clear; `/clear` and `/new` end the goal. The model
+  reports each turn via the goal-scoped `update_goal` tool
+  (`continue`/`complete`/`blocked`); report-less turns get one bounded
+  judge call when configured, else continue; unclear/failed judges pause
+  with the goal preserved. Three repeated tool results redirect with a
+  replan nudge (goal stays active). `complete` with unverified code or open
+  todos continues instead of stopping; `blocked` stops unconditionally
+  (declared-unverifiable checks print in the verdict, never gate). The live
+  goal rides every session save with stats intact (restore on `/resume` and
+  session switches; corrupt data loads as no goal); compaction appends a
+  `Goal:` line (text, state, stats, open todos) to the summary
+- Goal surface: status line shows `goal: <objective> [active|paused]`
+  (lowest-priority segment, hidden with no goal; `src/ui/status-bar.tsx`,
+  wired in `src/App.tsx`); turn telemetry traces carry the goal snapshot
+  with dashboard fragments and a Goal-turns card rendered only when goal
+  turns exist (`src/telemetry.ts`, `src/telemetry-dashboard.ts`); `/help`
+  lists `/goal` with subcommand descriptions; user docs in
+  `documentation/goals.md` (linked from `cli.md`, `index.md`,
+  `observability.md`), glossary entries in `CONTEXT.md`. Known limits: the
+  `update_goal` schema is not in the chat-payload tools list (the model
+  discovers it via continuation prose); multi-turn behavior against live
+  models is unproven (mocked suites only)
+
+- History caps removed (Pi parity): the 100-message / 200K-char ceilings are
+  gone — no `MAX_HISTORY_*` constants, no `ATOM_MAX_HISTORY_*` env vars, no
+  `maxHistory*` config keys (present keys are ignored as unknown), no trim
+  step in the loop, submit, resume, or session-switch paths. Full history
+  rides every POST; auto-compact at ~83% of the verified window plus manual
+  `/compact` is the only pressure valve. `LoopStats.truncationNotices` and
+  the 4-stage submit pipeline's budget-check stage are removed with them
+- Real context accounting (`src/adapters.ts`): Anthropic `input_tokens`
+  excludes `cache_read`/`cache_creation`, which previously made P%, NK, and
+  the 83% auto-compact trigger blind to cached context. Exclusive cache
+  counters are now folded into `prompt_tokens` at parse time (detail fields
+  stay provider-faithful), so load, spend, and compaction all see true
+  input-side tokens. OpenAI/Gemini paths already include cache — untouched,
+  with a regression test pinning no double-counting
+- Compaction retention raised (`src/compact.ts`): verbatim newest tail
+  8000 → 20000 estimated tokens (Pi parity); summary template restructured
+  to Objective / Important Details / Work State (Completed, Active,
+  Blocked) / Next Move / Relevant Files
+- Parallel-by-default reads (`src/scheduler.ts`): the conservative
+  same-tool+same-target overlap rule is gone — pure reads batch
+  unconditionally (same file, same task polls included). Same-file
+  read/write order, `bash` isolation, todo exclusivity, prompts, and
+  ordered commits are unchanged
+- Extension host (`src/extensions.ts`, `src/extension-commands.ts`,
+  `src/extension-ui.ts`, `src/project-trust.ts`, `src/tools/custom.ts`,
+  `src/tools/intercept.ts`, `src/tools/overrides.ts`,
+  `src/tools/provider-hooks.ts`, `src/tools/compaction-hooks.ts`):
+  project plus global extension scopes with trust-gated project execution,
+  enable/disable patterns, `--no-extensions` lockdown, model-callable custom
+  tools, pre/post tool interception, audited builtin overrides, slash
+  commands, session lifecycle events, provider request/response hooks,
+  compaction hooks, and status/widget/dialog UI. Guide plus three working
+  samples in `documentation/extensions.md` and `examples/extensions/`.
+  Known limits: non-UI registrations are runtime-global (no per-extension
+  unload); `session_start` covers startup/resume/switch/new; `overflow`
+  compaction reason is reserved for future callers; staged notices cap at
+  100 drop-oldest
+- `/effort` now applies on every provider (reasoning effort for
+  OpenAI-chat kinds, thinking budgets for Anthropic, thinking levels for
+  Gemini; `Auto` omits the knob) — previously zen-only
+- `reasoningEffort: default` renamed to `auto` (`default` still accepted as
+  an alias); `atom.example.json` gains an `extensions` example
+
 ## 1.2.0 — 2026-09-10
 
 - Durable multi-session store (`src/sessions.ts`): one JSON record per

package/README.md

@@ -30,6 +30,8 @@ Full docs live in [`documentation/`](documentation/index.md), same layout as ope
 - [Permissions and Modes](documentation/permissions.md) — normal/yolo/plan, trust, allow/deny rules
 - [Skills](documentation/skills.md) — discovery, frontmatter contract, auto-invoke
 - [Sessions](documentation/sessions.md) — durable multi-session store, rename, switcher, resume, clear, rewind
+- [Goals](documentation/goals.md) — pin one session objective that runs turn-to-turn
+- [Extensions](documentation/extensions.md) — zero-to-running guide plus working samples
 - [Observability](documentation/observability.md) — local telemetry, `/telemetry`, dashboard drill-down
 - [Compaction and Token Display](documentation/compaction.md) — auto-compact, manual compact, footer format
 - [Configuration](documentation/configuration.md) — env vars, auth file, AGENTS.md layering
@@ -113,13 +115,18 @@ in `package.json`, add a `CHANGELOG.md` entry, commit, tag `vX.Y.Z`, push —
   `~/.atom/auth.json`), `/effort` (reasoning-effort picker), `/tools`,
   `/skills`, `/skill:name` (invoke a skill; skills complete in `/`), `/context`
   (context usage by source), `/queue` + `/steer <text>` (follow-ups while
-  busy: queue until the turn ends, or inject into the running turn), `/help`, `/mode`, `/trust`,
-  `/clear`, `/exit` — plus `/`-autocomplete as you type (`/yolo` and `/plan` are retired as typed commands — `Tab` switches modes)
+  busy: queue until the turn ends, or inject into the running turn),
+  `/goal <objective>` (pin one session objective; bare shows it,
+  `pause`/`resume`/`clear` manage it), `/compact`, `/telemetry`,
+  `/dashboard`, `/rewind`, `/session`, `/resume`, `/rename`, `/new`, `/help`, `/mode`, `/trust`,
+  `/clear`, `/exit` — plus `/`-autocomplete as you type (`/yolo` and `/plan` are retired as typed commands — `Tab` switches modes). Full list: [CLI and TUI](documentation/cli.md)
 - 📊 **Status line** — provider · model · session token usage (`token:
   (P%) NK`: NK is the cumulative spend in K, P% is the current context load
   over the model's verified window — last `prompt_tokens`, else the
   4ch/token estimate; bare `token: NK` where no window is verified,
   `token: n/a` until reported — never estimated) · reasoning · mode, plus
+  `goal: <objective> [active|paused]` while a goal is live (lowest priority,
+  yields first under width pressure), plus
   live phase/elapsed/waiting while busy. It is the sole info bar: there is
   no persistent header, only the launch-time banner art.
 - 🗜️ **Context compaction** — auto-compacts at ~83% of the verified window
@@ -183,8 +190,10 @@ provider, chat. Switching provider keeps session history text; system prompt
 stays. `/model` is a unified picker: the active provider's live models first
 (fallback on any failure), then every other keyed provider's models plus the
 always-visible keyless Kilo list (free models carry a `(free)` badge) —
-picking one switches provider too. `/effort` sends `reasoning_effort` only
-for opencode-zen supported models; elsewhere kept but never sent. Your
+picking one switches provider too. `/effort` (`Auto`/`Low`/`Medium`/`High`/`Max`)
+applies on every provider — `reasoning_effort` for OpenAI-chat kinds,
+thinking budgets for Anthropic, thinking levels for Gemini (`Auto` omits
+it). Your
 `/model` + `/provider` + `/effort` picks persist across restarts (fresh
 conversation each launch; `/resume` restores it). Project defaults live in
 `atom.json` — see [Configuration](documentation/configuration.md).

package/atom.example.json

@@ -0,0 +1,11 @@
+{
+  "$comment": "ATOM config — copy to ./atom.json (project) or ~/.atom/atom.json (global). Every key is optional; unknown keys are ignored, invalid values fall back with a warning in /context. Precedence: env vars > saved session picks (/model, /provider, /effort) > project atom.json > global atom.json > compiled defaults.",
+  "provider": "kilo",
+  "model": "kilo-auto/free",
+  "reasoningEffort": "auto",
+  "maxToolSteps": 30,
+  "compactPct": 83,
+  "network": { "allowPublic": true, "allowLocalhost": true, "allowPrivate": false, "allowLinkLocal": false },
+  "telemetry": { "enabled": true },
+  "extensions": { "enabled": [], "disabled": [] }
+}