@xingular/harness 1.3.1-unstable.8 → 1.3.1-unstable.9

package/CHANGELOG.md

@@ -2,6 +2,87 @@
 
 ## Unreleased
 
+- **A completion check replaces `finish` (ADR-0071).**
+  - **The check.** When a reply that did work ends with its todo list closed or absent, the run's
+    own model, at low effort and with a fresh context, judges the person's request against what the
+    reply did: `done`, `continue` (the reply is sent back with the reason and next step), or
+    `needs_person`. The request is the goal; there is no `/goal` command.
+  - **Shared cap.** It shares the stop gate's cap of three nudges per reply and its no-progress rule.
+  - **Usage and failure.** Its request counts in the run's usage, and any failure leaves the reply
+    as it was.
+  - **`finish` is removed.** A live GPT 6 Luna run used `needs_decision` after 23 seconds to ask
+    something the handoff already answered. Only the check can now decide that the person is needed.
+  - **Prompt.** The `role` paragraph names which decisions are the agent's own, says a document the
+    request points to is the request, and says a question the request or files already answer is not
+    a question for the person. Its budget rises from 1,500 to 2,200 characters, so it fits uncut with
+    the subagent paragraph.
+  - **Kill switches.** `XINGULAR_STOP_GATE=off` and `XINGULAR_COMPLETION_CHECK=off` at launch switch a
+    layer off, for experiments.
+- **The model is told it has `edit` and `write`.** Pi's `createEditTool` and `createWriteTool` drop
+  the definition's prompt line and guidelines, and Pi lists only tools with a prompt line, so every
+  terminal and headless run's system prompt named only `read` and `bash`. Models edited files
+  through shell scripts instead: one 2-hour terminal run made 156 script edits (14 failed) and no
+  `edit` call. `createWorkspaceFileTools` now restores the prompt line and guidelines, as Pi's own
+  `createBashTool` does, for both the terminal and the runner. The stable prefix changes once.
+- **A plan no longer leaks its tools into normal mode.** A terminal plan request narrows the active
+  tools, and Pi records the restored normal set only with the next request. A rebuild before then
+  (a new runtime, resume, switch, fork or tree navigation) replayed the plan set, so normal work ran
+  without `edit`, `write`, browser and connected-app tools. One recorded session implemented a whole
+  plan that way. The terminal now marks each plan request in the session and restores the normal
+  set after any rebuild.
+- **Terminal replies continue across work slices (ADR-0070).** A terminal reply that reaches its
+  320-call or 60-minute budget after doing work now continues in place as the next slice, with a
+  checkpoint and a notice, up to 8 slices, instead of waiting for "continue" and re-sending the
+  whole context uncached. Fruitless, blocked, failed, aborted and plan slices still stop.
+  `XINGULAR_WORK_SLICE` also accepts key=value pairs such as `maxToolCalls=15,maxSlices=3`.
+- **One copy of each credential: no more `invalid_grant` from terminals.** A terminal's Pi kept the
+  account's tokens in `auth.json` under its own lock while the host, its subagents and usage probes
+  refreshed the same token in the Keychain under another, and `terminalEnd` copied the file back
+  over whatever the Keychain held. OpenAI rotates refresh tokens, so whichever side refreshed second
+  failed with `invalid_grant` and the account needed a new sign-in. A terminal now reads and refreshes
+  through the host store (`accounts.credentials.*`, INV-ACC-5), holding its lock for the refresh. An
+  `auth.json` left by an older terminal is folded in, newest token first.
+- **A stalled terminal reply is recorded as `stalled`.** Pi's abort reports back through
+  `terminal.interrupt` like a person's Esc, which closed the run as `user_abort` before the reply
+  settled as `stalled`; one silent 15-minute hang was recorded as the person's abort. The interrupt
+  now carries the stall. An interrupt also ends a settling reply's wait for its subagents.
+
+- **A reply can no longer quietly stop with work open (ADR-0069).**
+  - **Stop gate.** In attended top-level runs, terminal included, a reply that ends while its todo
+    list has open items is sent back to the model, naming the items, up to three times per reply,
+    and never after a nudge that produced no work.
+  - **`finish`.** Added here as the one explicit way to stop with work open; removed again by ADR-0071
+    above.
+  - **Honest stop codes.** Such a reply settles `interrupted` with the new stop codes `needs_person`
+    or `unfinished`, never `completed`, with a checkpoint the harness writes from the todo list,
+    files and checks.
+  - **`checkpoint_write` is removed.** A run used it as a sign-off and stopped with work open. At a
+    work-slice budget the harness now writes the checkpoint itself.
+  - **One budget per reply.** The work-slice budget now covers the whole reply: Pi's own
+    continuations (nudges, delivered subagent results, plan repair) no longer reset it.
+  - **Prompt wording.** The texts that read as permission to stop were rewritten: the todo guidance's
+    "final update before your final reply", and `## What would make this complete` "when the run
+    could be more complete".
+  - **Subagents.** `spawn_subagent` now returns the child's tools, and the read-only profiles say up
+    front when a brief needs a tool they lack.
+  - **Hosts.** Accept the two new stop codes and offer Continue for them.
+  - **Restart required.** A live terminal session keeps the code it started with, so restart it to
+    get the gate.
+- Do the job the request implies, and report it apart from the request. A standard session's role
+  paragraph now sets the scope of the work — the obvious adjacent step, what the work itself exposes,
+  the check it needs, inside the working folder and away from what a person must decide (a deploy, a
+  secret, remote data, money), without asking first inside those bounds — and the reporting guidance
+  requires two headings when they apply: `## Beyond what you asked` and
+  `## What would make this complete` (ADR-0068). The reporting budget rises from 2,000 to 2,400
+  characters so both fit uncut. All of it is stable-prefix text, so the cached prefix changes once
+  per session.
+- Say in the prompt when a reply may end. The standard session's role paragraph, the opening of the
+  reporting guidance and the todo guidance now say that an open todo item, a passing check or a
+  partial result is not a reason to end the turn, and name the reasons that are: the request is done,
+  the work-slice guard hands the reply back, a person's decision is needed, or the person interrupts.
+  ADR-0069 below now enforces it in the runner. A sweep of 822 runs in seven days found 207 that ended with work left
+  (137 within 30 tool calls, median 17, and 191 with no checkpoint), which is what the wording is
+  for. All three texts are in the stable prefix, so this changes the cached prefix once.
 - Reword the final-reply guidance so the outcome leads and structure follows what the answer needs. The
   reporting guidance is part of every request's stable prefix, so this changes the cached prefix.
 - Keep the provider cache working after a compaction (ADR-0067). The compaction checkpoint now arms only