altimate-code 0.8.2 → 0.8.4

package/CHANGELOG.md

@@ -5,6 +5,24 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.8.4] - 2026-06-05
+
+A trace-durability patch. Open `/traces` mid-session and you'd see a rich waterfall — then the moment the agent finished its turn the view collapsed to a single "system-prompt" span, the Summary tab's *"What was asked"* showed *"No prompt recorded"*, and the Chat tab dropped every user turn but the last. The data was genuinely gone from disk, not just hidden in the viewer. This release stops the on-disk trace from being overwritten after each turn and makes the file authoritative across worker restarts. A five-persona pre-release review drove a follow-up wording fix so a reconstructed trace isn't misread as a failed run.
+
+### Fixed
+
+- **Session traces no longer lose their data after every agent turn — the waterfall, summary prompt, and chat tab all stay intact.** A `session.status === "idle"` handler in the shared TUI worker fired once *per turn* (not once per session), ending the `Trace` and evicting it from cache; the next event reconstructed a fresh `Trace` whose near-empty initial state overwrote the rich `ses_<id>.json` on disk with a one-span file. Symptoms: the **Waterfall** collapsed to a lone "system-prompt" span, the **Summary** tab's *"What was asked"* went to *"No prompt recorded"*, and the **Chat** tab kept only the last user turn. The destructive idle handler is removed (sessions are long-lived — finalization happens on shutdown and `MAX_TRACES` eviction only), trace reconstruction now **rehydrates from the on-disk file** before falling back to a fresh start, and each user prompt is recorded as its own `user-message` span so multi-turn sessions render every turn in order. Authored-text scoping keeps synthetic parts (MCP banners, decoded file contents, reminders) out of the prompt and chat surfaces. Distinct from the v0.8.1 trace-corruption fix (#865), which addressed intra-instance concurrency — this is the worker-level cache lifecycle. (#895)
+- **A trace reconstructed after a restart no longer reads as a failed run.** When the on-disk trace is rehydrated (worker restart or `MAX_TRACES` eviction), any generation span still in flight is closed and marked so its boundary stays visible in the waterfall. The status message now states plainly that altimate-code restarted before the step finished recording and that this is **not an agent failure** — so an on-call reader debugging from a trace isn't sent chasing a phantom incident. (#895)
+
+## [0.8.3] - 2026-06-04
+
+A plan-mode reliability patch for the hosted gateway and other non-Anthropic models. Ask the `plan` agent to plan something benign — *"plan a feature to add a verify-output button"* — on `altimate-backend/altimate-default` (GPT-5.x) and it could refuse outright with *"I'm sorry, but I cannot assist with that request."* This release fixes the refusal at its source, hardens the fix against prompt-injection, and rewrites a warning that was misdiagnosing the symptom.
+
+### Fixed
+
+- **Plan mode no longer refuses benign requests on `altimate-default` and other non-Anthropic models.** The root cause was *how* altimate-code delivers its own plan instructions: a `<system-reminder>`-wrapped block attached as user-role text. Claude is trained to read that tag as authoritative system guidance, but GPT-5.x, Gemini, and other non-Anthropic models pattern-matched the same block — carrying `STRICTLY FORBIDDEN` / `ZERO exceptions` / `MUST` language — as a **prompt-injection attempt** and declined, returning *"I'm sorry, but I cannot assist with that request."* for ordinary developer tasks. altimate-code now **hoists those self-injected reminders into proper `role:"system"` messages** for non-Anthropic models, so they arrive as instructions rather than suspect user input; Anthropic models are unchanged. The hoist is scoped by **provenance** to altimate-code's own reminder parts — user-supplied files, MCP-resource bodies, and `data:`-URL content can never gain system-role priority, even when they begin with `<system-reminder>`. Gateway models are routed by `model.family` (via a shared `familyVendor` classifier) so a Claude- or Gemini-backed gateway model lands on the right base prompt instead of an Anthropic-style fallback. (#887, #888)
+- **The "plan agent used no tools" warning is accurate — and no longer false-fires.** Previously the warning blamed the model's *tool-use capability* (misleading — these models are fully tool-capable) and could fire on the final text-only step of a **successful** multi-step plan session, making a working session look broken. The trip-wire now scans the conversation history for prior tool calls, so a session that already explored the codebase isn't flagged. When it does fire, the copy describes the observed symptom only ("stopped without calling any tools"), lists the likely causes, and offers concrete recoveries — ask it to investigate first, rephrase, or, if it keeps refusing, `/model` to a tier more eager to explore. The plan agent is also now instructed to **read or search the codebase before drafting** any plan, with an explicit escape hatch for trivial, fully-specified changes. See [Plan mode](https://docs.altimate.sh/data-engineering/agent-modes/#plan) and [Troubleshooting](https://docs.altimate.sh/reference/troubleshooting/). (#888)
+
 ## [0.8.2] - 2026-06-03
 
 A small correctness patch. The dbt PR reviewer's deterministic engine never required an altimate API key — but two tool descriptions said it did, which could push the reviewer into lint-only mode and send users chasing a key they don't need. This release corrects that and documents what "lint-only" actually means.

package/package.json

@@ -14,20 +14,20 @@
   "scripts": {
     "postinstall": "bun ./postinstall.mjs || node ./postinstall.mjs"
   },
-  "version": "0.8.2",
+  "version": "0.8.4",
   "license": "MIT",
   "dependencies": {
     "@altimateai/altimate-core": "0.4.0"
   },
   "optionalDependencies": {
-    "@altimateai/altimate-code-linux-arm64": "0.8.2",
-    "@altimateai/altimate-code-darwin-arm64": "0.8.2",
-    "@altimateai/altimate-code-linux-x64-baseline": "0.8.2",
-    "@altimateai/altimate-code-darwin-x64-baseline": "0.8.2",
-    "@altimateai/altimate-code-windows-x64-baseline": "0.8.2",
-    "@altimateai/altimate-code-darwin-x64": "0.8.2",
-    "@altimateai/altimate-code-windows-x64": "0.8.2",
-    "@altimateai/altimate-code-linux-x64": "0.8.2"
+    "@altimateai/altimate-code-linux-arm64": "0.8.4",
+    "@altimateai/altimate-code-darwin-arm64": "0.8.4",
+    "@altimateai/altimate-code-linux-x64-baseline": "0.8.4",
+    "@altimateai/altimate-code-darwin-x64-baseline": "0.8.4",
+    "@altimateai/altimate-code-windows-x64-baseline": "0.8.4",
+    "@altimateai/altimate-code-darwin-x64": "0.8.4",
+    "@altimateai/altimate-code-windows-x64": "0.8.4",
+    "@altimateai/altimate-code-linux-x64": "0.8.4"
   },
   "peerDependencies": {
     "pg": ">=8",