# Agent Release Gate > Run the agent release gate — a portable, wire-level QA harness for the agent runtime. Drives the same product endpoint the playground drives and asserts on the SSE frame stream and real side effects, never on model prose, so it works against any deployment (cloud or self-hosted) from three env vars. Use before an agent-workflows release, or after changing the runner, the SDK agent adapters, the runner Docker images, or the agent service. Triggers: "run the release gate", "QA the agent runtime", "does the agent still work end to end", "pre-release agent QA". - Skill: `agenta-ai/agent-release-gate` (Agent Skill, multi-file: 54 files) - Install (CLI): `npx skillmds@latest add agenta-ai/agent-release-gate` - Raw SKILL.md: https://api.skillmd.com/api/skills/agenta-ai/agent-release-gate/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: DevOps & Infra - Author: Agenta-AI (https://skillmd.com/u/agenta-ai) - Updated: 2026-09-10 - Page: https://skillmd.com/skills/agenta-ai/agent-release-gate --- # Agent release gate Product-level sanity QA for the agent runtime, one layer below the playground UI. The question is not "is every detail right" — it is "**if a user opens the product and does the obvious first things, do they work?**" This is the gate a release passes before shipping. Every check asserts on the **wire** (the SSE frame types the browser sees) and on **side effects** (the file really persisted, the revision really incremented) — never on what the model says. That makes it deployment-agnostic: point it at any stack and the assertions still hold. ## Run it Set three environment variables for the deployment under test, then run the gate: ```bash export AGENTA_BASE=https://your-stack.example.com # deployment origin export AGENTA_PROJECT_ID=... # target project export AGENTA_API_KEY=... # project API key uv run resources/qa_product.py --all --custom-slug --custom-name "" --require-store # everything uv run resources/qa_product.py --cell P1 # one cell uv run resources/qa_product.py --cell C1 --only chat # one journey uv run resources/qa_product.py --cell S2 --only warm --only cold1 --require-store # continuity ``` **EXPORT the three variables, do not just set them.** The driver falls back to an env FILE when `AGENTA_*` is absent from its environment, which is helpful interactively and dangerous in a release run: a credentials file of bare `KEY=value` lines sourced with `. file` sets the shell only, the child `uv run` process inherits nothing, and the driver silently runs the whole gate against WHATEVER DEPLOYMENT the fallback file names. The failure surfaces as `401 Invalid credentials` from a stage whose key you just watched answer 200, or worse as a green run recorded against the wrong stack. Use `set -a` around the source, or `export` each variable, and confirm the stage in the results before trusting them. (Cost a staging gate run on 2026-08-28; the fallback degrades to "wrong deployment", never to "no credentials".) Paths are relative to this skill's directory. The deployment's vault must hold the provider keys the cells use (Anthropic / OpenAI / OpenRouter). If the three env vars are unset the driver stops immediately and names exactly what is missing; a legacy `--env-file ` fallback also exists. `--all` includes cells P2, P2b, and P3 (a custom OpenAI-compatible provider, with P2 and P2b running locally and P3 on Daytona), which need a vault slug passed via `--custom-slug`; the driver fails fast if it's missing. `--custom-name` is also required because `model_keys` is built from the display name rather than the stable slug. Custom-provider cells skip the credential-rotation journey because their value is write-only and cannot be safely restored. Cells S1, S2 and C1 additionally need the subscription sidecar logged in on the target deployment — see `resources/coverage.md` for what each cell requires. The Daytona cells (C2, C4, P3, X2) additionally run the `secret_opaque` journey and need the runner's Daytona API key to manage Secrets, because credential hiding is on by default; without it those cells fail at sandbox creation with an error naming the permission. **The one flag a release conductor must not skip past.** The continuity journeys (`warm`, `cold1`, `cold2`) only mean anything on a **store-backed** deployment: with no object store the runner degrades silently to an ephemeral working directory, so those journeys SKIP by default and FAIL with `--require-store`. Run the gate against a deployment with `AGENTA_STORE_*` configured and pass `--require-store`, or the greenest possible run still says nothing about durability. `cold2` additionally needs an operator hook that SIGKILLs the runner replica (`--cold2-replace-cmd`) and SKIPs without it. **The flag that makes the gate fit the release: `--release-base`.** The matrix is fixed, so without it a release that reworked a subsystem gets exactly the coverage of a release that did not touch it. Pass the ref the release branches from, and the driver reads the release's own changed paths, matches them against the rules in `resources/path_triggers.py`, and makes the cells those rules name MANDATORY for this run: ```bash uv run resources/qa_product.py --all --release-base origin/main --require-store # every release run uv run resources/path_triggers.py --release-base origin/main # preview only, runs nothing ``` A mandatory cell that lives in `qa_product.py` is added to the run even when `--cell` did not ask for it. A mandatory cell that is a standalone `matrix_*.py` script is a separate process the driver cannot observe, so it is printed, written to `mandatory.json`, and listed in `summary.md` under "Mandatory for this release" — **the release is not green until each of those has a recorded result of its own.** If a rule names a cell that does not exist, the driver stops before running anything and says so: the release changed code a rule protects and the coverage was never written, which is the one outcome that must never read as green. Add `--changed-path` to state paths by hand where the checkout is not the release branch. With no `--release-base` and no `--changed-path` nothing changes, so every existing invocation behaves exactly as before. Adding a rule is one line in `PATH_TRIGGERS` (a glob, and the cells it makes mandatory) plus the cell it names. Matching is `fnmatch` over the whole repo-relative path, so `*` crosses directory separators and `a/b/*` covers the whole subtree; write `**` so a subtree rule reads as one. **Reading the result.** Each journey prints `PASS`, `FAIL`, or `SKIP` with a one-line reason, and a per-cell markdown table lands with the full JSON in `./qa-gate-runs//` (override the location with `AGENTA_QA_RUNS_DIR`). Runs are written to the current working directory, never into the skill. `SKIP` is expected where a journey does not apply to a cell (for example `mcp` on any Pi cell — user MCP is Claude-only). Any `FAIL` blocks the release until triaged. **A SKIPped integration test in a security- or concurrency-bearing area is a FAILURE in your summary line, not green.** State it as "N passed, M skipped OF WHICH k are untested claims" and name the k. A commit-lock race test skipping for want of a reachable Postgres is exactly how a one-line syntax error (`SET LOCAL lock_timeout` with a bind parameter, which Postgres rejects outright) survived 1911 green tests before a human hit it as his first live action. **The two journeys that run many things at once: `burst` and `crosstalk`.** Every other journey drives one run at a time, so the gate only ever saw faults that reproduce on a quiet deployment. The credential-delivery fault of AGE-4249 does not: about one production first message in five failed because some fresh Daytona sandboxes start without their Secret substitution wiring, and per cold sandbox that is roughly an 8 percent fault. `burst` sends 16 first messages at the same time on 16 brand new sessions, so the run buys 16 cold starts instead of one. `crosstalk` runs 3 two-turn conversations with long output beside 2 approval flows, and checks that no stream carries another session's nonce, except on the codex harness, where the gate rides a platform tool with empty arguments, so the approval command carries no nonce and isolation is not checked there (`nonce_checked=false`). Both are Daytona-only by default and skip elsewhere; both report the runner's stable error code per run, so a `credential_delivery_failed` names itself. ```bash uv run resources/qa_product.py --cell C4 --only burst --only crosstalk uv run resources/qa_product.py --cell C4 --only burst --burst-size 24 # more cold starts uv run resources/qa_product.py --cell C3 --only crosstalk --concurrency-everywhere # local too uv run resources/test_qa_product_concurrency.py # offline tests ``` **Read a green burst honestly.** At an 8 percent per-cold-start fault rate, 8 runs miss the fault 51 percent of the time, 16 miss it 26 percent, and two Daytona cells at 16 miss it about 7 percent. A PASS is a sample, not an all-clear, and the result says so in its own reason line. A FAIL is proof. Each concurrent run holds its own Daytona sandbox, about 5 GiB of the organization's disk, and a parked sandbox keeps counting until its auto-delete window closes, so a burst of 16 is about 80 GiB in flight. The counts are `--burst-size` (default 16), `--crosstalk-conversations` (default 3) and `--crosstalk-approvals` (default 2). The cap is 32 concurrent runs: 32 for the burst size, and 32 for the two crosstalk counts TOGETHER, because what costs disk is what runs at once. When the provider refuses on capacity the journey reports SKIP with a loud reason, never a PASS or a FAIL, because nothing about the product was measured. `--concurrency-timeout` (default 300s) bounds one TURN and rides into the stream as an absolute deadline, so a two-turn crosstalk run gets twice that and a stream that never ends is abandoned rather than followed. A release that changes `services/runner/src/engines/sandbox_agent/**` or `services/runner/src/providers/daytona*` makes the Daytona cells C2, C4 and X2 mandatory through `path_triggers.py`, and forces `burst` and `crosstalk` into the run even when `--only` named something else. That is how these journeys reach a release that needs them. **Before a human gets a deployment URL, run `resources/qa_commit_approval.py` too.** It is not part of `qa_product.py`'s cell × journey matrix — none of that matrix's journeys drive a live turn against a REAL, saved workflow revision (the `commit` journey only exercises the REST API; `chat`, `tool`, `approve`/`deny` all run against an inline, unsaved config), so nothing else in the gate observes the S3b single-use execution-authorization gate actually firing around a real config mutation. This script does: create a real workflow + revision, invoke a live agent turn that calls `read_config` then `commit_revision`, expect the pause, approve it in-band, and verify the new revision landed with REST fetch-back. Treat a FAIL here as blocking, the same as any other gate FAIL. ## Tiers: coached vs. mechanism-blind Every cell in this gate (`qa_product.py`'s journeys, `qa_commit_approval.py`, `qa_probe.py`, and all `matrix_w*.py` cells) declares which of two tiers it belongs to, in its own docstring: - **Coached (backend-path test).** The prompt names the mechanism verbatim — which tool, which operation, which target path. This proves OUR CODE works (the gate, the base check, the authorization handoff, the sandbox path) when the right call is made. It proves NOTHING about whether a model finds that call from a plain-language human ask. - **Mechanism-blind (model-behavior test).** The prompt is phrased the way a real user types — no tool names, no operation names, no schema hints. Only these cells license a claim about what the model can do unprompted. **Rule: claims of model behavior may only cite mechanism-blind cells.** Most cells in this directory are coached tier — they test backend paths, correctly, but do not stand in for model-discovery evidence. The gap this rule exists to close, found live: asked in plain words to "add the skill I saved in your folder," Haiku invented a nonexistent marker syntax (`{"@ag.embed": {"@ag.references": ...}}`) and the engine accepted it as literal data — a failure none of the coached cells (including `matrix_w7.py`, whose prompt names `@ag.file` outright) could ever have caught, because they never test whether the model reaches for the real mechanism on its own. `resources/matrix_g1_guidance_discovery.py` is this directory's first mechanism-blind cell, promoted after the platform-guidance fix closed that exact gap; it reuses `qa_matrix_lib.py` the same way the separate one-shot benchmark (Tier B) does — check there before writing a new mechanism-blind cell from scratch, to avoid duplicating scaffolding. ## Session control cells `resources/session_control.py` is a second, standalone driver: sixteen cells that cover Stop, durable commands, and the runner's recovery paths (owner release, park/resume, watchdog quarantine). It drives the same product endpoint and asserts on the same wire, but it needs its own account bootstrap, so it runs as a separate process rather than as `qa_product.py` cells. See `resources/path_triggers.py` for the exact mandatory-cell mechanism. **These cells are MANDATORY** — run them, not just the standing gate — whenever the release diff touches any of: - `services/runner/src/sessions/**` - `services/runner/src/engines/sandbox_agent/**` - `api/oss/src/core/sessions/**` - `api/oss/src/tasks/asyncio/sessions/**` - `api/oss/src/apis/fastapi/sessions/**` Run every cell with one line: ```bash uv run resources/session_control.py --cells all --harness pi_core --sandbox local ``` Add `--project ` to run the eight cells that need direct Docker and Postgres access (`sandbox-gone`, `records-outage`, `restart-after-stop`, `runner-gone`, `runner-gone-late`, `post-stop-row`, `codex-child`, `stale-tail`) and the abort-log subcheck inside `stop-after-finish`. Without `--project`, those eight cells SKIP with a named reason. The `stop-after-finish` HTTP check still runs, but only its abort-log subcheck is unavailable. The other eight cells (`stop-warm`, `double-send`, `stale-stop`, `stop-approval`, `stop-after-finish`, `repeat-stop`, `concurrent-stops`, `stop-during-completion`) run over HTTP alone against any deployment. Add `--resume ` to pick a lost run back up: any cell already recorded there is loaded instead of re-run. Results land in a timestamped folder under `~/agenta-qa-evidence/` (override with `AGENTA_QA_RUNS_DIR`), as `results.json` and `summary.md` — the same PASS/FAIL/SKIP shape as the rest of the gate. When a release path makes session control mandatory, pass that artifact to the standing gate with `--session-control-results `: a missing or incomplete artifact stops the gate before the matrix runs, and a recorded FAIL makes the final gate exit nonzero. **Environment, by name.** Same three-variable discipline as the rest of the gate, no env-file fallback: - `AGENTA_BASE` — the deployment origin. - `AGENTA_ADMIN_KEY` — mints the ephemeral account this driver runs under. Lives in `~/.agenta-qa-secrets.env`. - `QA_OPENAI_API_KEY` — stocked into that account's vault so the `pi_core` and `codex` harnesses have a provider key. Lives in `~/.agenta-qa-openai.env`. - `ANTHROPIC_API_KEY` — only required for `--harness claude`, stocked into the same vault the same way. Lives in `~/.agenta-qa-secrets.env`. A pi_core- or codex-only run does not need it. A Daytona run additionally needs a Secrets-capable Daytona key on the runner; the key in most session env files returns 403 on the Secrets endpoint, so check that before trusting a Daytona result. ## When results lie The runtime **fails open**: a component can break, get logged, and the turn still succeeds with a normal-looking answer. A green turn is therefore not proof on its own. Before trusting a pass, read `resources/LESSONS.md` — every trap there produced a green test that proved nothing. The three that bite hardest: replay conversation history byte-faithfully (tool parts included) or every turn silently goes cold; re-run any prior blocker-level finding after a redeploy before believing it; and a multi-turn check that never leaves the warm daemon, on a deployment with no object store, proves nothing about the durable working directory (LESSONS #16). ## Resources (read on demand) - `resources/coverage.md` — the cells (harness × sandbox × auth), the journeys (chat, mount, tool, approve, deny, commit, warm, cold1, cold2, mcp) with a one-line meaning for each, the continuity tiers and their method, and a table of what each cell needs beyond the three env vars. - `resources/LESSONS.md` — the traps. Read before writing or trusting any agent QA test. - `resources/qa_product.py` — the gate driver (cells × journeys). - `resources/matrix_gw1_gateway_tools.py` — **[coached, with one mechanism-blind leg]** the gateway tool surface against a real provider: search filters by policy (the denied key never reaches the model), an allowed tool executes unattended with a genuine provider result, and an ask-tier tool parks with the right stored identity and is answered through the **interactions API** — the durable plane a reloaded browser uses, which no other cell covers. The fixed matrix proves approvals with a builtin, so nothing else notices when the compiled policy and the enforced policy drift apart. Defaults to the no-auth `text_to_pdf` connection; `--integration` and `--connection` point it elsewhere. SKIPs, naming the fixture, when no valid connection exists, and SKIPs rather than failing the release when the model provider itself errors. Made mandatory by the gateway rule in `path_triggers.py`. - `resources/path_triggers.py` — the path-scoped rules: one dict mapping a path glob to the cells a release must run when its diff touches that glob, plus the two functions the driver calls. Runs standalone as a preview (`--release-base `). Add a rule here whenever new coverage is only meaningful for changes in one part of the tree. - `resources/qa_probe.py` — a one-turn wire probe: `uv run resources/qa_probe.py` confirms the product path answers at all before running the full gate. - `resources/qa_commit_approval.py` — **[coached]** the mandatory pre-handoff commit-approval round trip (see above). Self-contained; does not import `qa_product.py`. - `resources/qa_matrix_lib.py` — shared helpers (session/turn plumbing, workflow/revision REST calls, the multi-round approval loop) for the `matrix_w*.py` adversarial cells below. Import only, no CLI. It also holds the two cross-cutting invariants every cell should fold into its verdict — `check_no_blank_success_on_refusal` and `check_no_silent_turn` (see below). **If you write a new cell, wire `check_no_silent_turn` into its PASS condition.** - `resources/matrix_w3.py` — **[coached, with a narrow mechanism-blind sliver]** two sessions, disjoint edits; session B is given a stale `base_revision_id` and ZERO coaching on recovery. Two-tier pass: autonomous correct recovery passes outright; a model that diagnoses the 409 correctly and asks before re-attempting a config write ALSO passes (that is desirable caution, not a failure) once one bare "yes, retry" permission (no mechanics) completes the recovery. Guards both the optimistic-concurrency base check and the instructive-error design (a 409 must be readable without coaching). The initial action in both sessions is still coached — only the RECOVERY step is mechanism-blind; don't cite this cell for "the model finds commit_revision unprompted." - `resources/matrix_w4.py` — **[coached]** a pending approval whose base goes stale while it waits (a second session commits first); the EXECUTE-time check must catch it, not just the gate-time check. - `resources/matrix_w5.py` — **[coached]** interrupt a running turn (steer) then use the session again. Caught a real bug: the durable mount never gets re-established after a steer, breaking every subsequent turn on that session. Distinct from Mahmoud's own steer repro (that one is a turn-currency/heartbeat bug; this is a mount-lifecycle bug) — keep the two separate when triaging. - `resources/matrix_w7.py` — **[coached]** an agent writes a workspace file and commits it via an `@ag.file` marker; asserts the approval manifest carries digest+bytes and the commit lands with the exact bytes. Caught a real bug (fixed, PR #5763-adjacent runner fix): the gate approved cleanly but execution always refused with `authorization_missing` — no file-marker commit could land. Re-verified PASS after the fix. Note: `DEFERRED_NOT_EXECUTED` on a queued second tool call is benign (another gate is already pending), not a real tool error — don't let it fail this cell. Naming `@ag.file` verbatim in the prompt is correct for this cell's backend-path purpose but means it cannot catch a model failing to find the marker syntax unprompted — see Tiers above. - `resources/matrix_w1_daytona.py` — **[coached]** the commit round-trip on `sandbox=daytona` (every other cell runs local). Needs a funded provider vault key (see the file's own docstring for the exact vault-secret shape and the `provider_key` slug gotcha). PASS confirms `DaytonaWorkspaceReader` and the placeholder-secrets flow live. - `resources/matrix_t8_saved_files.py` — **[coached]** T8: the Daytona remote agent mount (the durable `agent-files/` folder the playground file drawer writes into), which needs the ngrok tunnel the runner discovers at sandbox-acquire time. Writes a marker file into the mount via the mounts API (`POST /mounts/agents/sign` + `PUT /mounts/{id}/files`) before the run, then asserts the runner log line `remote agent mount active for artifact=` appears, a tool output actually carries the real marker content (not hallucinated), and the commit lands with it. Verified PASS 3/3 runs after the tunnel-seat fix (2026-08-06); this line was absent on every attempt before that fix landed. - `resources/matrix_t9_agent_tools.py` — **[coached]** T9: the runner restores the agent's own tools from `agent-files/.tools/` before a session (`agent-tools-setup.ts`). Plants a `setup.sh` and a `bin/qa-tool` through the mounts API, opens a fresh session, and asserts the first tool call is the exact probe and its output payload carries both planted tokens, plus a line `setup.sh` appended to `agent-files/.tools/runs.log`, read back through the mounts API with no model in the loop, plus the `agent_tools_setup` stage for THIS session in the runner log when `--runner-container` is given. `--sandbox local|daytona`, `--harness pi_core|claude`. Mandatory (via `path_triggers.py`) when the restore step or the sandbox image recipes change. - `resources/matrix_w7_per_harness.py` — **[coached]** matrix_w7.py's exact scenario run identically on all three harnesses (claude, codex, pi_core), each classified PASS/FAIL/SKIP independently. Exists because W7 originally ran on Claude only, and that scenario-coverage gap is exactly what let the Codex approve-then-fail P0 (2026-08-06) ship: commit_revision + file marker + HITL approval on a harness this suite never exercised. claude uses subscription auth (no vault dependency); codex and pi_core need a funded OpenAI `provider_key` vault secret (mirrors cells X1/C3 in `qa_product.py`) and correctly SKIP with the exact reason when it's missing or ambiguous — a SKIP here is an untested harness, not a pass, and must be named as such in any release summary. `--only ` runs a single leg without re-spending budget on the others. Verified PASS on claude (2026-08-06). codex/pi initially SKIPPED on this shared preview stack because its vault held zero — then, transiently, an ambiguous multiple — OpenAI candidates (other concurrent agents' activity on the same shared project); resolved 2026-08-06 by stocking one unambiguous OpenAI `provider_key` secret (no stale entries existed to remove). Re-verified PASS on codex after stocking the key (session 58ce3a58-8d04-40ac-99e4-c44eaa5d7b06). - `resources/matrix_w7_daytona.py` — **[coached]** matrix_w7.py's exact scenario with `sandbox=daytona` instead of local. The local-only original W7 is exactly why this bug hid: the Daytona transport rejects NUL bytes in argv, which the `@ag.file` manifest walk was emitting, so no workspace-file commit could EVER land on Daytona, on any harness, until the 2026-08-06 fix (found during the same P0 triage, live-verified twice on codex+Daytona: sessions f3fa4335, f2f22056). This cell is the sandbox-axis regression guard, staying on the claude harness. Needs the same funded Anthropic vault key as `matrix_w1_daytona.py`. Verified PASS (2026-08-06). - `resources/matrix_invariant_commit_auth_refusal.py` — **[coached scenario; the invariant itself is mechanism-level]** the generic invariant: no `tool_result` with empty output and `isError:false` may exist for a call whose runner log says `[commit-auth] refused` (the silent-blank-success class — the P0's actual failure shape, distinct from the scenario-coverage gap `matrix_w7_per_harness.py` addresses). The check itself (`qa_matrix_lib.check_no_blank_success_on_refusal`) reads the runner's own log line against the wire outcome and is meant to be reusable by any cell that exercises marker-carrying commits, not just this one. This cell's trigger is best-effort: run W7's flow, let the legitimate commit consume its authorization record, then REPLAY the byte-identical approval-carrying request (a duplicate submission) hoping to force a second, doomed `authorizeExecution` attempt. Verified 2026-08-06: the replay did NOT reproduce a refusal (the runner treats the identical replayed history as already-resolved and answers conversationally instead of re-invoking the tool), so the cell correctly SKIPped rather than claiming a false pass on an invariant it never exercised. A reliable deterministic trigger (e.g. a genuine cold-resume stale-approval replay, or a crafted duplicate `toolCallId`) is an open follow-up; until then, treat any SKIP from this cell as "the invariant was not tested this run," never as green. - `qa_matrix_lib.check_no_silent_turn(turns)` — **[mechanism-level invariant; no cell of its own]** no turn may come back completely bare: no text, no tool call, no approval gate, no file or data payload, and no error. That combination is a swallowed provider failure (ASD-EST100) — the model call is rejected, the error is dropped on the way back, and the turn is reported as a clean empty finish, so the user sees a blank bubble with no reason anywhere. It matters most in cells whose PASS depends on something NOT appearing (no error, no leak, no blank success): a turn that produced nothing satisfies those by doing nothing at all. Its content definition deliberately mirrors `content_parts_emitted` in the product's own Vercel egress (`sdks/python/agenta/sdk/agents/adapters/vercel/stream.py`) — reasoning does NOT count, so a turn that only thought is still a violation. Wired into `matrix_w7*.py`, `matrix_t8_saved_files.py`, `matrix_b1_builtin_find.py`, `matrix_invariant_commit_auth_refusal.py`, `matrix_l3_abandoned_approval.py`, `matrix_w3.py`, `matrix_w4.py` and `matrix_w5.py` as `... and not silent["violations"]`; add the same conjunct to any new cell. The one thing a cell must exclude itself is a turn it deliberately aborted or interrupted, which legitimately ends bare — `matrix_w5.py` shows the pattern (it checks only its post-interrupt turns). - `resources/matrix_g1_guidance_discovery.py` — **[mechanism-blind]** does the platform guidance actually change what the model does? The trial prompt is Mahmoud's own verbatim phrasing from the live session that found the bug ("can you add gstack-autoplan skill to your skills (i saved it in your folder)") — no tool, operation, or marker syntax named. Before the guidance existed, this exact phrasing made a model copy the skill into its own harness-local skills folder and claim success without ever proposing a commit, 3/3 live (session b59cb549). Two parts per harness/model/sandbox leg: PROBE reads the rendered instructions file out of the workspace and asserts the fenced platform-guidance block and the skill-location sentence are really there; TRIALS run the prompt N times, PASS only when a commit_revision gate fires, is approved, and the STORED revision carries the skill (copying into the harness's own folder is a FAIL). Setup gotchas baked into `qa_matrix_lib.py` (`PI_CORE_HARNESS_KIND`, `PI_CORE_HAIKU_MODEL`): the Pi harness kind enum is `"pi_core"` (bare `"pi"` 500s), and `pi_core` rejects a bare `"haiku"` model id (needs `"claude-haiku-4-5"`); codex accepts its curated short alias (`"gpt-5.6-luna"`) bare. Both legs need `sandbox=daytona` + a vault key (codex: OpenAI; pi_core: the same Anthropic key `matrix_w1_daytona.py` documents) and SKIP with the exact reason when the credential is missing or ambiguous. `--only ` runs a single leg without re-spending trial budget on the other. Neither leg has yet scored a clean 4/4 live, and both failure modes are real, reproducible model mechanics misses, not noise or infra flake — treat this cell's discovery rate as a genuine open quality question, not a settled pass: - **pi_core: 2/3** (2026-08-06). Trial 2's model ran `cp -r agent-files/gstack-autoplan .agenta-imports/` (copying the whole directory) then referenced a marker path that didn't match where the file landed; the engine correctly denied it fail-closed (`approved-content resolution failed ...: gstack-autoplan/SKILL.md does not exist under .agenta-imports/.; deny`). Not a product bug — the deny is doing its job — but evidence the model fumbles the exact copy-then-reference mechanics some of the time. - **codex: 2/3** (2026-08-06, re-verified after stocking a clean OpenAI vault key — the earlier SKIP was purely the missing/ambiguous credential, now resolved). Trial 1 didn't attempt the mechanism at all: zero gates approved, the model just replied "I can't add skills by simply placing files in the skills folder; skills must be enabled through the agent configuration" (session 4fa17164-3ada-4ef6-86b5-63bf1c64f10b) — it named the right concept but never called read_config/commit_revision to act on it. Trials 2 and 3 passed cleanly (sessions 4c6a758b-39a3-4789-890e-06f3d68196b6, bf8a1283-667c-45c7-ab0d-b112e12106db). Worth a call: whether the guidance text needs to be more directive (e.g. explicitly say "copy the FILE, not the directory" and "always attempt the tool calls, don't just describe the mechanism"), or whether ~2/3 is an acceptable bar for this feature's launch. ### Builtin capability cells (`matrix_b*.py`) — does the harness's own tooling still work - `resources/matrix_b1_builtin_find.py` — **[coached, harness-mechanism test]** one native file-search call per harness (write three known marker files, then ask the model to locate them via its own search capability — never a manual directory listing), asserting the exact filenames come back in the TOOL OUTPUT payload, never the reply. Exists because nothing else in the gate ever exercised a harness builtin: verify-runner's overnight diagnosis found Pi's `find` builtin dead 52/52 across two benchmark runs (it shells out to the vendored `fd` binary with a flag that only exists from fd 9 onward; the runner image ships fd 8.6.0) — a total capability loss that sat invisible with nothing calling it. This cell closes that discoverability gap for the class, not just this one instance. **Open discrepancy, not yet reconciled**: this cell's own pi_core leg PASSED twice, live, with real filenames back (2026-08-07, sessions dd9c51ef-92d0-4780-855e-da6f48e07d9f and 47f02ab1-61ec-4db8-a0d6-2d96e98b9188) — which does not match "52/52 failed". Either the break is conditional on a flag/option this cell's simple case never exercises, or something already changed; needs reconciling with verify-runner before Pi's `find` gets called either fixed or still broken. **Codex SKIPs by design**, not tested: its exec output doesn't land in the `tool-output-available` payload's `.output` field (the same quirk `qa_product.py`'s `j2_mount` already names and skips codex for), so this cell's evidence extraction cannot see codex's real results — a codex-shaped extraction is a follow-up. Claude PASSED cleanly on the two-turn version (session 4081e9ee-11c1-4f12-8245-6c390f90e9d8). Needed two EXPLICIT turns (write, then search) — one combined instruction left claude stopping after the write step without attempting the search at all. ### The lifecycle cells (`matrix_l*.py`) — cold ↔ warm, and what survives each transition These four cover the session-lifecycle work: which config changes are applied to a RUNNING sandbox and which tear it down, and what happens to a pending approval, a client tool and the durable mount across each transition. They all assert the STORED turn ledger (`POST /sessions/turns/query` → one `sandbox_id` per turn) or the STORED interaction rows (`POST /sessions/interactions/query`), never the SSE echo — nothing about warm-versus-rebuilt ever reaches the stream. An empty ledger FAILS a cell; missing evidence is not evidence. - `resources/matrix_l1_lifecycle_routes.py` — **MANDATORY. [mechanism-blind]** the routing matrix itself: for each kind of mid-conversation config change, assert the route the runner took. One sandbox id = applied in place, two = rebuilt. Blocks on six cases: no change must stay warm; an instructions edit, a permissions edit and a tool-catalog edit must escalate; and a same-connection model switch must stay warm on BOTH claude and pi_core. The pi_core model case (added 2026-08-29) is the standing trap for the wire-spelling bug class: the router once keyed its table on the bare "pi" literal while the wire carries "pi_core", every playground model switch silently rebuilt, and the claude-only case could not see it (#6364). This is the cell that would have caught the `cold1` rot described below. - `resources/matrix_l2_approval_across_config_change.py` — **MANDATORY. [coached]** the killer combination: an approval answered while a config change rides along in the SAME request. It is the regression test for the applied-state bug (the pool used to stamp the INCOMING fingerprint on the approval-resume path, so the next turn continued warm on an environment running something else). Asserts the gated commit lands, the approval row ends `resolved`/`responded`, and — the real tell — the config change is not swallowed: it takes effect on the FOLLOWING turn, via a rebuild, in both the instructions and the permissions variant. - `resources/matrix_l3_abandoned_approval.py` — **MANDATORY. [coached]** the user sends a new message instead of answering the card. Asserts the gated tool does NOT run (an unanswered approval is not consent), the row is swept to `cancelled` rather than left `pending`, and the session still works. `cancelled` vs `pending` is the loud-vs-silent distinction: a `pending` row is a card sitting on the page that no process is waiting on. - `resources/matrix_l5_live_route_observed.py` — **MANDATORY. [mechanism-blind, with a control]** the other half of L1: an instructions edit made mid-conversation must actually be OBSERVED by the harness, not merely written to disk. Runs the same configuration on a fresh cold session as a control, so a failure isolates the runner rather than blaming the model; when the control also fails it reports INCONCLUSIVE instead of a confident wrong verdict. It asserts the edit, never the route, so it stays meaningful if the facet is ever made live again. **Failed 2026-08-06 (claude/local) and now passes** — see the finding note below. Extended 2026-08-06 (overnight gate run) to all three harnesses in one invocation (`--only ` for a single leg) — this MANDATORY blocker cell had only ever run on claude; codex and pi_core were a named gap. Verified PASS on all three the same night the harness matrix landed. - `resources/matrix_l4_client_tool_lifecycle.py` — **nice-to-have. [coached]** the client-tool round trip, and the only cell that covers client tools at all. Asserts the browser's result reaches the model and the `client_tool` interaction is stored. It RECORDS rather than asserts the sandbox count, which is two today: a client-tool pause is deliberately not parkable (`"warm-hold": RESERVED, not built`, #5384), so every client-tool round trip currently costs a rebuild. If that number ever reads one, the warm hold landed and the docstring needs updating. - `resources/matrix_n1_session_context.py` — **MANDATORY. [journey, with two controls]** the per-turn session facts on the path the PRODUCT uses. Renames a session between two turns and asks the agent for the name, asks the agent for its own display name, and posts a forged `meta.session_context` that the service must ignore. Every expected value carries a random token minted for the run and spoken nowhere in the conversation, so a transcript-derived answer cannot match; the second ask is the exact shape of #6661, because by then the FIRST name is in the transcript and the current one is only in the stored header. This is the one cell that asserts on model prose, and it does so because nothing else can: `turnContext` is a prompt string on the service-to-runner payload, the runner logs nothing, and it is deliberately kept out of `request.messages` and out of persisted input. Two controls keep a FAIL honest — an echo probe (a model that cannot repeat a literal token makes the cell report INCONCLUSIVE) and a read-back of the stored header after every rename. Mandatory (via `path_triggers.py`) when the SDK session-context module, the platform-prompt renderer, the agent handler, or the API's session-context resolver changes. **The gate could not see this class at all before v0.115.3**: #6661 shipped green because the API stamped the facts in its invoke prelude, which never runs for a playground turn, and no cell renamed a session mid-conversation or asserted on the facts. **Finding the lifecycle cells surfaced (2026-08-06, claude on local, reproduced 3×) — FIXED:** the `workspaceFiles` live route rewrote the instruction file and advanced applied state, but the running harness never re-read it. A warm session kept obeying the instructions it started with while the pool reported the NEW fingerprint, so every later turn matched and continued warm and the user's edit had no effect until something else evicted the session. A cold session with the identical configuration obeyed it immediately, which is what isolated the runner. This was the failure `desired-state.ts` refuses to allow for the `prompts` facet ("refreshing them and claiming the model saw the change would be a lie") reappearing on the facet that WAS made live — and note the direction: before the live route existed, an instructions edit forced a rebuild and therefore took effect on the next turn, so it was a regression in what the user sees, not a speedup. `matrix_l5_live_route_observed.py` is the repro. The fix withdrew the route: `workspaceFiles` now routes to `rebuild-sandbox` in the capability table, and `refresh-workspace` left `LIVE_ACTION_KINDS` so restoring the table alone fails closed. An instructions edit costs a sandbox again, which is what it cost before the optimisation. **L1's `instructions` case therefore expects TWO sandbox ids, and L2's first variant expects two as well** — if you are reading an old green from before 2026-08-06, those cells expected one. The intended next shape is refresh THEN reopen the session, which needs the reopen to build its session init from the incoming request first, and needs proving on L5 rather than asserting. **Why `cold1` changed (read before trusting an old green):** that tier used to force its eviction by editing `instructions.agents_md`, which the lifecycle work briefly made a LIVE route — the tier would have gone on passing while measuring warm reuse, and `park`'s "one sandbox id is meaningful" argument rests on `cold1` reporting two on the same deployment. It now moves `harness.permissions` (the `harnessSession` facet → `reopen-session`, deliberately not live) and ASSERTS two distinct sandbox ids. It sta …(truncated)