pi-session-mcp 0.8.3 → 0.12.0

package/docs/evaluation/s5-decision-inputs-2026-10-18/worker-usage.md

@@ -0,0 +1,63 @@
+# Research note: worker token usage on terminal turns (#92)
+
+Date: 2026-10-18. Source: main `612fbcb`, pinned `@earendil-works/pi-coding-agent`,
+`pi-agent-core` and `pi-ai` **1.2.0** (versions read from the installed
+`package.json` files after `npm ci`). Research only: no MCP field, no runtime
+change and no internal import follows from this note. Any new public field needs
+its own decision and a minor release under ADR 0003.
+
+A read-only Pi worker did the first pass; its line references came from a stale
+`node_modules` tree that still held 1.1.0 although the lockfile pins 1.2.0. The
+statements below were rechecked by the integrator against the 1.2.0 artifacts.
+Package-relative paths: **AI** = `pi-ai`, **CA** = `pi-coding-agent`,
+**AC** = `pi-agent-core`.
+
+## What the public root exports carry
+
+- `Usage` is reachable from the package roots (`exports["."]` only). It has
+  `input`, `output`, `cacheRead`, `cacheWrite`, `totalTokens` and
+  `cost.{input,output,cacheRead,cacheWrite,total}`, plus two optional fields
+  (AI `dist/types.d.ts:274-295`).
+- Assistant messages always carry `usage` (AI `dist/types.d.ts:370`;
+  CA `docs/message-types.md:69`). Tool results may carry `usage` for nested model
+  work, which "is not part of the main model-call usage" (CA
+  `docs/message-types.md:170-177`).
+- `compaction_end.result` can carry the summary call's `usage`
+  (CA `dist/core/agent-session.d.ts:73`, `dist/core/compaction/compaction.d.ts:24`,
+  `docs/compaction.md:153,165`). Automatic retries are separate session events
+  (`auto_retry_start` / `auto_retry_end`, CA `dist/core/agent-session.d.ts:80,86`).
+- `AgentSession.getSessionStats()` is cumulative over the whole session,
+  including compacted history (CA `dist/core/agent-session.d.ts:883-886`), so it
+  is not a per-turn value.
+- The adapter already subscribes to root-exported `AgentSessionEvent`s and keeps
+  the last assistant message, but reads no usage today (`src/sdk-pi-adapter.ts`).
+
+## Documented semantics versus inference
+
+| Topic | Documented | Not documented or inferred |
+| --- | --- | --- |
+| `reasoning` | Subset of `output`, already included; `undefined` when the provider gives no breakdown (AI `types.d.ts:281-286`; CA `docs/message-types.md:90`) | — |
+| `cacheWrite1h` | Subset of `cacheWrite` (same places) | — |
+| `input` vs `cacheRead` overlap, `totalTokens` formula | — | Not defined by the declarations; summing fields would risk double counting |
+| Scope of one record | One assistant response, i.e. one model request; streaming `usage` is the latest cumulative value **within that response** and may stay zero until completion (CA `docs/json.md:95`) | — |
+| One pi-session-mcp turn | — | Inferred: one `pi_session_prompt` turn spans several model requests (tool loop, AC `README.md:72-105`), possibly retries and compaction; no documented per-prompt aggregate exists |
+| Missing vs zero | — | Not specified provider-wide; a reported 0 cannot be told apart from "not reported" |
+| `cost` | Catalog price of the model (AI `types.d.ts:493`) | Not an invoice and not subscription cost |
+
+## Consequence for a possible proposal
+
+A terminal turn could carry optional usage only if it is named precisely, for
+example "sum of the `usage` of the assistant responses observed in this turn",
+with each field summed separately, `reasoning` never added to `output`,
+`cacheWrite1h` never added to `cacheWrite`, nested tool usage and compaction
+usage reported separately or excluded, and `null` when no assistant response was
+observed. Whether that sum equals what the provider bills is **not** established,
+and retries, compaction and nested tool calls make "per turn" a derived figure,
+not an SDK-provided one.
+
+Unsupported SDK need, recorded rather than worked around: a documented per-prompt
+usage aggregate (covering retries, compaction and nested tool work) and a
+documented distinction between "not reported" and zero.
+
+Next step only with a separate decision: a proposal with exact field semantics,
+null rules and a minor release. Nothing in this sprint implements it.

package/docs/evaluation/s7-live-acceptance-2026-10-24/README.md

@@ -0,0 +1,212 @@
+# Sprint 7 live acceptance of the #118 candidate (2026-10-24)
+
+Part of [Sprint 7](../../sprints/2026-10-24-s7-stabilization-live-acceptance.md).
+Every result below is bound to the candidate, not to the published 0.11.2.
+
+## Source binding
+
+| Item | Value |
+| --- | --- |
+| Candidate commit | `6651a47096ad515c8075b374fd5775ee965e6067` (branch `fix/118-zombie-thread-group`), clean detached checkout |
+| Package version string | `0.11.2`, unchanged on purpose; the candidate is identified by commit and digests |
+| `dist/main.js` | `eebfe6ff59cfadec385ce90b58e6083c236cc83b5f18115c3368f20b80151fc4` |
+| `dist/` tree (kit digest, 123 files) | `00ce08b967edf57671a7f7d75aa6dd57ef543897914eb301a699dc51ec603102` |
+| `package-lock.json` | `a1390d69f83444889e9688ab0ea477fe3ec03746d908c0b20c933714a1e6c7db` |
+| Packed tarball | `f897e11509b69e62397c93e022163010c9dd97bf4ee487075b28e15492e160bb` |
+| Difference to the published 0.11.2 `dist/` | only `setup-platform.js` and `setup-platform.d.ts` (122 files compared); the wake path (`wake-main`, turn events, hook) is byte-identical. Runtime dependency difference: `zod` 4.6.7 instead of 4.6.5 |
+| Pi SDK | 1.4.0, all family packages 1.4.0 |
+| Clients | Claude Code 2.1.308, Codex CLI 0.162.1 |
+| Runtime | Node.js 26.11.1, npm 12.2.1, Podman 6.1.4 (rootless, runc), Linux 7.2.10 x86-64 |
+| Controller models | Claude Code `claude-sonnet-5-5` (pinned); Codex built-in default under `--ignore-user-config` (not exposed by `codex exec --json`) |
+| Worker routes | read-only `explore-fast` (openai, gpt-6-luna, low); coding `code-deepseek` (deepseek, deepseek-flash, high) |
+| Live limits (maintainer) | 20 worker starts, 32 worker turns, 120 minutes, one targeted rerun; no money limit was set |
+| Used | 20 worker starts, 23 worker turns, about 48 minutes; the one rerun was used (W2r) |
+
+The kit (fixtures, oracles, driver, leak scan) was prepared per cycle outside
+the repository, as `docs/acceptance.md` prescribes for Level 3. The checksums
+of its scripts and prompts are in [`evidence/kit.sha256`](evidence/kit.sha256);
+the pre-registration was fixed before the first model turn, with an addendum
+also fixed before that turn (E3b, E6b, isolation mode, budgets).
+
+## Level 1
+
+On the candidate checkout: `npm ci`, `npm run typecheck`, `npm run build:cli`,
+`NODE_OPTIONS=--unhandled-rejections=strict npm test` (114 files, 3348 passed,
+13 skipped), `npm run build`, `git diff --check` and an empty `git status`
+all exit 0.
+
+**`npm audit` fails:** 1 high (`undici` 8.10.2, pinned exactly by Pi SDK
+1.4.0, [#120](https://github.com/plagor44-ux/pi-session-mcp/issues/120)) and 1
+moderate (the SDK package as its dependent). The same run on the earlier
+candidate `e80679e` also reported `zod` (high) and `fast-uri` (moderate); this
+sprint patched those. Level 1 is therefore not green.
+
+## Level 2
+
+`scripts/setup-acceptance.mjs`, all four exit 0 with zero failures, real
+registrations unchanged, no leftover process, no path leak, disposable homes
+removed:
+
+| Run | Checks |
+| --- | --- |
+| Working tree, delivery off | 160/160 |
+| Working tree, `--events spool` | 228/228 |
+| Packed tarball, delivery off | 160/160 |
+| Packed tarball, `--events spool` | 228/228 |
+
+Files: `evidence/l2-*.json`.
+
+## Level 3 (ten tasks, two clients)
+
+Strictly sequential, three invocations per client, per-invocation MCP
+definitions (`--strict-mcp-config` for Claude Code, `-c` overrides for Codex),
+delivery off, synthetic fixtures with baseline digests, hidden acceptance
+tests outside the fixtures, a snapshot barrier before every coding follow-up.
+
+| Task | Claude Code | Codex |
+| --- | --- | --- |
+| 1 Capability-led read-only analysis | passed | passed |
+| 2 Same-session read-only follow-up | passed | passed |
+| 3 Bounded coding change | passed | passed |
+| 4 Same-session coding follow-up | passed | passed |
+| 5 Abort and recovery | passed | passed |
+
+10 of 10 tasks reached their intended outcome; every oracle check passed.
+Leak scan: 0 hits, no undocumented error code. Read-only fixtures unchanged.
+No candidate process and no session left after any invocation (checked at
+exit and 10 s later). No `LATE.txt` or leftover `slow.mjs` 135 s after each
+abort. Real MCP registrations and the production hook count unchanged.
+
+Polling (descriptive; both clients could `sleep`): `pi_turn_get` calls per
+turn 1–3, running polls 0–1.
+
+**Gate:** every row of the [Level 3 gate](../../acceptance.md#gate) holds
+except "Level 1 is green for the release source", which fails on
+`npm audit` (#120). The cycle is therefore **not accepted as a release
+gate**; its ten task outcomes stand as recorded.
+
+Preflight P0 (two routes, PONG) failed once on a kit defect before any worker
+started: the runbook pointed the preflight at the Level 3 configuration, whose
+aliases differ, and the server answered `unknown_workspace` as documented. The
+corrected preflight passed. Both attempts are in `evidence/` (`p0a-*` is the
+failed one).
+
+## Event delivery
+
+Claude Code runs in stream-json sessions held open by the driver,
+`--setting-sources project,local` (the production user-scope hook does not
+take part), a private `XDG_RUNTIME_DIR`. Setup ran with a disposable home, so
+ownership state stayed disposable.
+
+| Scenario | Outcome | Observed |
+| --- | --- | --- |
+| E1 setup path, one turn | passed | Setup `--dry-run`, `--apply` (`applied`, `skill_installed`, `hook_added`) and `--verify` as expected; hook exactly once in the project settings. The waiter exited 2 at 16.8 s; a new controller request started 1.6 s later while idle and read the turn once; no poll, no sleep |
+| E2 two concurrent turns | passed | Wakes at 22.5 s (short) and 86.1 s (long), each its own controller request about 2.1 s later; each turn read once; 0 running polls; both check ids correct |
+| E3b budget wake (manual hook: 40 s budget, 100 s timeout) | passed | Budget wakes at 45.5 s and 87.5 s, each followed by exactly one read that returned `running` and re-armed the hook; terminal wake at 107.8 s, one terminal read; turn 102 s; id correct. A manual hook, not evidence for the setup path |
+| E4 abort and close | passed (see note) | Abort `ok`; exactly one waiter exit 2 (A, `aborted`); closing the running B woke nothing (its waiter exited 0); A read once, B never; list empty; nothing left after 135 s |
+| E5 Codex without `XDG_RUNTIME_DIR` | passed | The server environment had no runtime directory; capabilities `off`; one inactive warning, no startup failure; turn completed by polling; no wake claimed |
+| E6b hook pickup, running stream-json session (CLI) | picked up | The hook was installed by setup 6 s after the session started; the next `pi_session_prompt` armed it and the wake started a new request 1.4 s after the waiter exited |
+
+E4 note: the kit's oracle also checked that the wake started its own
+controller request. Here the wake arrived while the controller was still in
+its own round (closing B) and was delivered with that round's tool result.
+The pre-registered E4 criteria do not require an own request, so the verdict
+follows them; the oracle's extra check is recorded as failed in
+`evidence/e4-oracle.json`.
+
+### VS Code extension (E6)
+
+The candidate was not loaded into a VS Code session: that would have replaced
+a production registration, which this sprint may not do. Evidence on the
+reference build only: the integrator's own Claude Code session in the VS Code
+extension, with the published 0.11.2 and its setup-installed user-scope hook,
+received two wakes while idle. Each started a new controller round that read
+the turn once without polling (worker turns of the #118 fix and of its review,
+10:03–10:22 UTC). The candidate's wake path is byte-identical to that build.
+Not checked in VS Code: whether a running extension session picks up a newly
+installed hook. Shortest operator check: open a project with no hook in a
+VS Code window, start Claude Code there, run `pi-session-mcp-setup --apply
+--target claude-code:local:pi-session-mcp` from that project in a terminal,
+then ask the running session to start a worker turn and end its turn; a wake
+without a restart or `/hooks` review answers the question.
+
+## Container writer with delivery
+
+A writer created only for this sprint: synthetic repository, alias `s7wfix`,
+profile `code-deepseek`, `events.delivery` spool, installed from the packed
+candidate tarball, real home (writer state, a project-local Claude Code
+registration and ownership records are created there by design and were
+removed).
+
+| Step | Outcome |
+| --- | --- |
+| Dry runs, isolation control | planned; control `isolation_control_passed` |
+| Provisioning, first apply | `writer_provisioned`, then **`writer_acceptance_failed`** |
+| Provisioning, second apply | `writer_accepted` (`pass`, network `not_limited`), `writer_registered`, `hook_added` |
+| `--writer --verify` | `mcp_verified`, `hook_verified` |
+| W2 coding delegation by Claude Code | passed: turn 72 s, one terminal wake at 78.4 s, own request 2.2 s later, one read, no poll; worker committed `multiply` on a new branch |
+| Collect, review, finish | `writer_collected`; the imported ref changes only `math.js` and `math.test.mjs`; an independent check and `node --test` on an export of the ref pass; `writer_reset` |
+| W2r (the one rerun) | the delegation passed again the same way; collect, review and finish passed |
+| Remove | `writer_removed`, `hook_removed`; registration gone; ownership file byte-identical to before; no container, process or writer state left |
+
+The first acceptance failure could not be diagnosed: setup keeps only the
+code and removes the harness report. The harness, run by hand with the same
+arguments, passed twice. Tracked as
+[#121](https://github.com/plagor44-ux/pi-session-mcp/issues/121).
+
+Boundary checks, provider-free, with `podman inspect` and `podman exec` in a
+container started through the writer's own wrapper:
+
+- Mounts: clone (rw), clone hooks and Git config (ro), contained config (ro),
+  agent copy (rw), `/pi/events` (rw, the writer's own `events/`), release (ro).
+- The writers root shows only the own writer, and that only `clone`; a decoy
+  namespace `s7decoy-w9` with a canary was not reachable (0 finds). Of the
+  host home only the path scaffold is visible; `.ssh`, the runtime directory
+  and the ownership file are not.
+- Writes to `/`, `/pi`, the release and the contained config fail
+  (read-only file system); `/pi/events` is one mount.
+
+Crafted events in the writer's live instance directory (server mode, written
+from inside the container), against the candidate's `wake-main` with a
+synthetic running turn:
+
+| Input | Hook result |
+| --- | --- |
+| Nothing | budget wake (exit 2, budget text) |
+| Junk name, invalid JSON, 2000-byte file, symlinked event file, extra key, unknown state, event in a subdirectory, symlinked "instance directory" | ignored; only the budget wake |
+| Well-formed `via: close` for the turn | silent at once (no wake for a closed session) |
+| Well-formed forged `completed` for the turn | terminal wake (exit 2) telling the controller to read the turn once: a spurious wake, as documented |
+| A second waiter for that claimed event | silent (no double delivery) |
+| Next launch | removed the leftover files and the symlink |
+
+**Not run:** the forged event during a real controller turn. In W2 and W2r
+the kit's oracle ran `podman exec` with the system runtime directory, while
+the client had started Podman with the private one; rootless Podman keeps its
+runtime state there, so `exec` found no container and nothing was placed.
+Shown afterwards provider-free (the same `exec` succeeds with the client's
+runtime directory). The rerun allowance and the start limit were used up, so
+it was not repeated. The controller rule it would test, that a wake is only a
+trigger and the read state decides, was exercised in E3b, where each budget
+wake led to one read that returned `running` and to further waiting.
+
+## Cleanup
+
+All sessions started by this sprint were closed. No candidate server, waiter,
+worker or setup process, no container of the test writer, no spool instance
+directory, no hook entry written by this sprint and no writer state remain.
+The production MCP registrations, the production hook count, the controller
+skills in the homes and the ownership file are unchanged. The test run
+directory, the disposable homes and the installed candidate prefixes were
+removed after the evidence was copied. Claude Code keeps its own project
+entries for the scratch directories in `~/.claude.json` (no MCP servers);
+`claude purge` can drop them and was left to the operator.
+
+## Limits
+
+- Ten small synthetic tasks and targeted wake scenarios on one host, one pair
+  of client versions and two worker routes; not model capability on larger
+  work.
+- No token, cache or money figures: Claude Code reports cost only on the API
+  key route, and Codex exposes no per-request count. No saving is claimed.
+- Wake reaction times are single observations.
+- macOS, Codex wake and the TUI were not part of this cycle.

package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/binding.json

@@ -0,0 +1,20 @@
+{
+  "schema": 1,
+  "packageName": "pi-session-mcp",
+  "packageVersion": "0.11.2",
+  "distMainSha256": "eebfe6ff59cfadec385ce90b58e6083c236cc83b5f18115c3368f20b80151fc4",
+  "distTreeSha256": "00ce08b967edf57671a7f7d75aa6dd57ef543897914eb301a699dc51ec603102",
+  "distFiles": 123,
+  "versions": {
+    "node": "v26.11.1",
+    "npm": "12.2.1",
+    "claude": "2.1.308 (Claude Code)",
+    "codex": "codex-cli 0.162.1"
+  },
+  "platform": "linux-x64",
+  "source": {
+    "commit": "6651a47096ad515c8075b374fd5775ee965e6067",
+    "cleanTree": true,
+    "tag": null
+  }
+}

package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/config-ev-off.json

@@ -0,0 +1,43 @@
+{
+  "schema": 1,
+  "events": "off",
+  "configSha256": "a037db554cb6ceae8d008bbd3843b379937a506d57ed5d5b3ec0aa79c867e0b1",
+  "workspaceAliases": [
+    "ev-e1",
+    "ev-e5",
+    "ev-e2s",
+    "ev-e2l",
+    "ev-e3",
+    "ev-e3c",
+    "ev-e4a",
+    "ev-e4b"
+  ],
+  "workspaceModes": {
+    "ev-e1": "read-only",
+    "ev-e5": "read-only",
+    "ev-e2s": "host",
+    "ev-e2l": "host",
+    "ev-e3": "host",
+    "ev-e3c": "host",
+    "ev-e4a": "host",
+    "ev-e4b": "host"
+  },
+  "executionProfiles": [
+    {
+      "alias": "explore-fast",
+      "default": true,
+      "permissionProfile": "read-only",
+      "provider": "openai",
+      "model": "gpt-6-luna",
+      "thinkingLevel": "low"
+    },
+    {
+      "alias": "code-deepseek",
+      "default": false,
+      "permissionProfile": "coding",
+      "provider": "deepseek",
+      "model": "deepseek-flash",
+      "thinkingLevel": "high"
+    }
+  ]
+}

package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/config-ev-spool.json

@@ -0,0 +1,43 @@
+{
+  "schema": 1,
+  "events": "spool",
+  "configSha256": "b3a04aaee93fcbd879c4b6bb98eaf4ebd64314af8900820f7babcad57e501771",
+  "workspaceAliases": [
+    "ev-e1",
+    "ev-e5",
+    "ev-e2s",
+    "ev-e2l",
+    "ev-e3",
+    "ev-e3c",
+    "ev-e4a",
+    "ev-e4b"
+  ],
+  "workspaceModes": {
+    "ev-e1": "read-only",
+    "ev-e5": "read-only",
+    "ev-e2s": "host",
+    "ev-e2l": "host",
+    "ev-e3": "host",
+    "ev-e3c": "host",
+    "ev-e4a": "host",
+    "ev-e4b": "host"
+  },
+  "executionProfiles": [
+    {
+      "alias": "explore-fast",
+      "default": true,
+      "permissionProfile": "read-only",
+      "provider": "openai",
+      "model": "gpt-6-luna",
+      "thinkingLevel": "low"
+    },
+    {
+      "alias": "code-deepseek",
+      "default": false,
+      "permissionProfile": "coding",
+      "provider": "deepseek",
+      "model": "deepseek-flash",
+      "thinkingLevel": "high"
+    }
+  ]
+}

package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/config-l3.json

@@ -0,0 +1,39 @@
+{
+  "schema": 1,
+  "events": "off",
+  "configSha256": "24597fdaa550ff84b69c5ae0bdb651c2e647a4cff741f5a05c0add4ee3c929cb",
+  "workspaceAliases": [
+    "cc-analysis",
+    "cc-code",
+    "cc-abort",
+    "cx-analysis",
+    "cx-code",
+    "cx-abort"
+  ],
+  "workspaceModes": {
+    "cc-analysis": "read-only",
+    "cc-code": "host",
+    "cc-abort": "host",
+    "cx-analysis": "read-only",
+    "cx-code": "host",
+    "cx-abort": "host"
+  },
+  "executionProfiles": [
+    {
+      "alias": "explore-fast",
+      "default": true,
+      "permissionProfile": "read-only",
+      "provider": "openai",
+      "model": "gpt-6-luna",
+      "thinkingLevel": "low"
+    },
+    {
+      "alias": "code-deepseek",
+      "default": false,
+      "permissionProfile": "coding",
+      "provider": "deepseek",
+      "model": "deepseek-flash",
+      "thinkingLevel": "high"
+    }
+  ]
+}

package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/config-pre.json

@@ -0,0 +1,31 @@
+{
+  "schema": 1,
+  "events": "off",
+  "configSha256": "773d8d451a8f2c7b693ff3cb6b49e503e91082f6057255b12da6f79fb78a24d7",
+  "workspaceAliases": [
+    "pre-analysis",
+    "pre-code"
+  ],
+  "workspaceModes": {
+    "pre-analysis": "read-only",
+    "pre-code": "host"
+  },
+  "executionProfiles": [
+    {
+      "alias": "explore-fast",
+      "default": true,
+      "permissionProfile": "read-only",
+      "provider": "openai",
+      "model": "gpt-6-luna",
+      "thinkingLevel": "low"
+    },
+    {
+      "alias": "code-deepseek",
+      "default": false,
+      "permissionProfile": "coding",
+      "provider": "deepseek",
+      "model": "deepseek-flash",
+      "thinkingLevel": "high"
+    }
+  ]
+}

package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e-teardown-setup-remove.json

@@ -0,0 +1 @@
+{"schemaVersion":1,"operation":"remove","status":"ok","findings":[{"target":{"client":"claude-code","scope":"project","alias":"pi-session-mcp"},"status":"ok","code":"removed"},{"target":{"client":"claude-code","scope":"project","alias":"pi-session-orchestration"},"status":"ok","code":"skill_removed"},{"target":{"client":"claude-code","scope":"project","alias":"hook:pi-session-mcp-wake"},"status":"ok","code":"hook_removed"}],"exitCode":0}

package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e1-clean-10s.json

@@ -0,0 +1,15 @@
+{
+  "schema": 1,
+  "label": "e1+10s",
+  "clean": true,
+  "counts": {
+    "candidateServers": 0,
+    "wakeWaiters": 0,
+    "setupProcesses": 0,
+    "workerCommands": 0,
+    "clientProcesses": 0,
+    "spoolInstances": 0,
+    "waiterLocks": 0
+  },
+  "claimFilesInformational": 1
+}

package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e1-clean.json

@@ -0,0 +1,15 @@
+{
+  "schema": 1,
+  "label": "e1",
+  "clean": true,
+  "counts": {
+    "candidateServers": 0,
+    "wakeWaiters": 0,
+    "setupProcesses": 0,
+    "workerCommands": 0,
+    "clientProcesses": 0,
+    "spoolInstances": 0,
+    "waiterLocks": 0
+  },
+  "claimFilesInformational": 1
+}

package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e1-digest-ev-e1.json

@@ -0,0 +1,8 @@
+{
+  "identical": true,
+  "baselineTreeSha256": "ed7793f666b6f8da5a51aad3d1bc8f72076114280c673bf2a6dc524c68e4cbba",
+  "currentTreeSha256": "ed7793f666b6f8da5a51aad3d1bc8f72076114280c673bf2a6dc524c68e4cbba",
+  "added": [],
+  "removed": [],
+  "changed": []
+}

package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e1-leaks.json

@@ -0,0 +1,131 @@
+{
+  "schema": 1,
+  "transcriptSha256": "7aa9bc42ace670368f17f8c7d2b46d3ef6a94ef10cf335a235c4fa6c43ce3e96",
+  "transcriptName": "transcript.jsonl",
+  "format": "claude",
+  "lines": 30,
+  "parseErrors": 0,
+  "hasTimes": true,
+  "scannedToolResults": 6,
+  "hitCount": 0,
+  "hits": [],
+  "assistantTextMentions": {},
+  "calls": {
+    "total": 6,
+    "byTool": {
+      "pi_capabilities_get": 1,
+      "pi_session_start": 1,
+      "pi_session_prompt": 1,
+      "pi_turn_get": 1,
+      "pi_session_close": 1,
+      "pi_session_list": 1
+    },
+    "piTurnGetTotal": 1,
+    "piSessionCalls": 4
+  },
+  "errorCodes": {},
+  "undocumentedErrorCodes": [],
+  "perTurn": [
+    {
+      "session": "s1",
+      "turn": "t1",
+      "piTurnGetCalls": 1,
+      "runningPolls": 0,
+      "terminalReads": 1,
+      "finalState": "completed",
+      "durationMs": 10634
+    }
+  ],
+  "perSession": [
+    {
+      "session": "s1",
+      "workspace": "ev-e1",
+      "profile": "explore-fast",
+      "closed": true,
+      "aborts": []
+    }
+  ],
+  "controller": {
+    "modelRequests": 7,
+    "resultEvents": 2,
+    "bashToolUses": 0,
+    "toolUses": {
+      "Skill": 1,
+      "pi_capabilities_get": 1,
+      "pi_session_start": 1,
+      "pi_session_prompt": 1,
+      "pi_turn_get": 1,
+      "pi_session_close": 1,
+      "pi_session_list": 1
+    },
+    "costUsdLast": 0.2157596
+  },
+  "wakes": {
+    "distinct": 1,
+    "occurrences": 2,
+    "rows": [
+      {
+        "line": 15,
+        "tMs": 16810,
+        "eventType": "system",
+        "subtype": "hook_response",
+        "kind": "terminal",
+        "turn": "t1",
+        "state": "completed"
+      },
+      {
+        "line": 15,
+        "tMs": 16810,
+        "eventType": "system",
+        "subtype": "hook_response",
+        "kind": "terminal",
+        "turn": "t1",
+        "state": "completed"
+      }
+    ]
+  },
+  "wakeRounds": [
+    {
+      "waiterExitLine": 15,
+      "waiterExitMs": 16810,
+      "requestLine": 19,
+      "requestMs": 18397,
+      "delayMs": 1587,
+      "delivery": "own_request"
+    }
+  ],
+  "hookEvents": [
+    {
+      "line": 11,
+      "tMs": 6151,
+      "subtype": "hook_started",
+      "hookEvent": "PostToolUse",
+      "outcome": null,
+      "exitCode": null
+    },
+    {
+      "line": 15,
+      "tMs": 16810,
+      "subtype": "hook_response",
+      "hookEvent": "PostToolUse",
+      "outcome": "error",
+      "exitCode": 2
+    },
+    {
+      "line": 21,
+      "tMs": 19159,
+      "subtype": "hook_started",
+      "hookEvent": "PostToolUse",
+      "outcome": null,
+      "exitCode": null
+    },
+    {
+      "line": 23,
+      "tMs": 19219,
+      "subtype": "hook_response",
+      "hookEvent": "PostToolUse",
+      "outcome": "success",
+      "exitCode": 0
+    }
+  ]
+}