pi-session-mcp 0.8.3 → 0.12.0

package/docs/evaluation/token-footprint-2026-10-17/follow-up.md

@@ -0,0 +1,76 @@
+# Local follow-up proposals; nothing published or approved
+
+*Status 2026-10-18:* the M1 pilot passed on its third attempt
+([m1-pilot](m1-pilot/README.md)). Sprint 5 implemented candidate 1 as skill
+guidance (#86), candidate 2 as worker-skill guidance (#87), and a slimmer input
+schema (#90); their live effect is unmeasured. The A/B for them is proposed,
+unapproved, in [`../m1-ab-proposal-2026-10-18/`](../m1-ab-proposal-2026-10-18/README.md).
+Statements below about a missing pilot decision are historical.
+
+The current evidence is linked from `README.md`; the first calibration remains
+separately preserved in the parent JSON records. Independent native review and
+verification attempts are recorded alongside each binding. Maintenance continues; isolation (#79) retains
+priority. This file authorizes no implementation, provider request or publication.
+
+| Priority / candidate | Observed benefit and quality | Cost / compatibility / decision |
+| --- | --- | --- |
+| 1. Consume a terminal result once | Two labelled re-fetches add 47,830 response bytes and two calls; removing them preserves T1 in 5/5 cases. No consistent latency gain. The M1 pilot's L3 repeated four terminal fetches live (one required). | Small controller-ledger/skill practice, current eight tools, no ADR/release change required for an operator prompt convention. Recommended for the tested lifecycle, with no claim that current real usage repeats results. |
+| 2. Compact worker reports | Removes 22,056 response bytes per tested terminal read/repair, 5/5 independent successes in each arm. Failed/blocked report fields retained in ten controls; no live obedience or larger-review proof. | Small draft skill instruction; conditionally recommended for a later approved A/B. Keep status, actual changes, checks including failures, blockers, risks and relative evidence. Load full logs/diffs only when necessary. 4 KiB is a test candidate, not a universal cap. |
+| 3. Discovery/schema reduction | Fresh discovery is 20,009 wire bytes; array 19,917. The synthetic Codex whole request tools array is 21,412 bytes; real-route reinjection/cache behavior and token impact remain unknown. | Not established. First measure concrete client projection/usage. Preserve discovery compatibility; any public field/tool change needs its own decision and minor release under ADR 0003. |
+| 4. Remove one representation | Two equal server representations; the synthetic Codex request probe sees one selected marker occurrence for normal, error and truncated replies. | No duplicate-removal benefit established for the tested client route. Preserve compatibility fields. Cloud routes and other clients remain unmeasured; a public change needs a separate decision and minor release. |
+| 5. Longer pauses / deterministic external waiting / pi_turn_wait | Current live need unmeasured; long M0 timings are deliberately synthetic. | Existing pause/ledger instructions suffice for now. No wait tool justified under ADR 0003. A supported concrete client wait path must be measured before proposing it. |
+| 6. Validated server compact mode / extensions / worker thinking off | No current live bottleneck or token benefit demonstrated. Thinking is already excluded as structured public output. | Not established; reject production implementation in this session. Internal worker usage is a different target. No compressor/cache/framework installation. |
+
+## Prepared development path and smallest next decision
+
+A dedicated `--footprint-live` adapter now represents L1/L2/L3 beside the existing
+profile runner. It preserves `validateApproval` and adds a separate exact-schedule
+operator decision binding current sources, compiled build, controller skill,
+manifest, budget and routes. The generic runner and its schemas remain intact.
+The three execution templates are generated by `--footprint-plan`; all approval
+flags remain false and the advisory amount remains null. They are proposals,
+not authorized model use.
+
+The existing gate requires a positive **operator-selected advisory** USD amount
+and explicit acknowledgment of unmetered use. That amount is neither a measured
+subscription cost nor a hard ceiling. No value was invented here. Once the
+operator supplies the exact pilot decision and advisory amount, bind the approved
+manifest/budget/decision digests and execute exactly three fresh sequential Codex
+controllers, four worker starts and five work turns, without whole-run retries.
+No historical approval is reusable.
+
+L3 uses two independent read sessions sharing the same disposable immutable T1
+fixture, then one follow-up in the first session. Re-fetching that original turn
+for invariance is explicit verification overhead, separate from normal terminal
+consumption. Each turn is keyed by actual session/turn identity transiently;
+records export only ordinals and allowlisted metrics. Independent oracles and
+scope snapshots remain separate from worker claims.
+
+The first pilot has since passed (see `m1-pilot/README.md`). Only after it establishes usable
+routes, client-event measurements and cleanup should a new, specifically approved
+three-pair alternating A/B compare compact reports or terminal-consumption practice.
+No live quality, token saving, model ranking or Level-3 claim follows from the
+provider-free calibration. The local model-request fixture establishes only its
+declared Codex/loopback route. No provider configuration, client registration,
+credential store or installed skill was changed by this development work.
+
+## Draft skill instruction for a later approved A/B
+
+Request a concise report containing task status, actual changed paths, checks with
+exit codes and failures, blockers, risks and relative evidence references. Include
+source-linked findings needed to verify the task. Keep full diffs/logs out of the
+default report and retrieve them when needed. Preserve failures even when the turn
+completed. A size target is advisory: if required evidence will not fit, report the
+limit explicitly instead of dropping it. The controller verifies files/tests and
+consumes each terminal result once. This was a local draft on 2026-10-17; no skill
+was installed or changed then. Sprint 5 later implemented it in the worker skills (#87).
+
+## Public-contract follow-up, only if later evidence warrants it
+
+Title: Evaluate an explicit compact-result mode after measured client token benefit.
+Entry gate: three approved paired runs show independent quality retention and a
+material observed controller-usage improvement, with actual client projection,
+latency, failed checks, malformed/truncated reports and compatibility measured.
+Decision must specify existing/default versus opt-in behavior, truth/privacy
+limits, required fields, rejection behavior and a minor release. Current evidence
+does not meet that gate; no ADR is accepted or external issue opened here.

package/docs/evaluation/token-footprint-2026-10-17/integrity.json

@@ -0,0 +1,171 @@
+{
+  "schema_version": "token-footprint-integrity-1",
+  "source_digest": "sha256:33c450601727e38cf63b1874df0fcadc158c3d06af3b472b3d1603da6383c929",
+  "artifacts": [
+    {
+      "path": "README.md",
+      "sha256": "sha256:2a12800b933c46788fca39a1b894a287a90ff0ddf0784d674216c62804ba59c5",
+      "bytes": 17879
+    },
+    {
+      "path": "baseline.json",
+      "sha256": "sha256:96975cfc685b4f041682acddf8fbe6b6fab642473ca9e58d651b19e7658f1605",
+      "bytes": 784197
+    },
+    {
+      "path": "calibration-2/README.md",
+      "sha256": "sha256:983acc6974c03c44d84ee64d571d60e35db51d8c136c1ec51a0edc8b0573cf62",
+      "bytes": 1553
+    },
+    {
+      "path": "calibration-2/baseline.json",
+      "sha256": "sha256:178e19b1dae5411d9b7fbc71c41e0fded649e54d5f14c2cee2ba620aa8d9f477",
+      "bytes": 917106
+    },
+    {
+      "path": "calibration-2/client-projection.json",
+      "sha256": "sha256:164be9fe4a9d9d3798f24bb302eb594765f23fbc4739899222e1a5df99cb9158",
+      "bytes": 80400
+    },
+    {
+      "path": "calibration-2/live-approval/budget.json",
+      "sha256": "sha256:a765198f0a7cefa2ae210166eb92a07e4edad2139023fae75c6bdeb629ae04dd",
+      "bytes": 290
+    },
+    {
+      "path": "calibration-2/live-approval/decision.json",
+      "sha256": "sha256:fbe2491603b6dc2618204c8352ecadf4ebdeb19a7eeafba21ecb01cc6951fae8",
+      "bytes": 1602
+    },
+    {
+      "path": "calibration-2/live-approval/manifest.json",
+      "sha256": "sha256:b52a9c76cdfc7a9e27de5f947a40153aba0fda9766de6a8abec3bd4b0c23282c",
+      "bytes": 8473
+    },
+    {
+      "path": "calibration-2/live-package.json",
+      "sha256": "sha256:ab2f7ff6b3c9fcb1f001e2ef72517e150c52fae3a2fbd4f6d3a7f02e39e3e271",
+      "bytes": 24912
+    },
+    {
+      "path": "calibration-2/measurement.json",
+      "sha256": "sha256:aaf24558aa4a93ca417ae0e8e370e52d6db403b7f04e79eb09908113f9190185",
+      "bytes": 1142216
+    },
+    {
+      "path": "calibration-2/summary.json",
+      "sha256": "sha256:65b3dd321e97ac50263630a14e1c0218fa21dd9f3233347bdfb3536c219ab4c5",
+      "bytes": 3380
+    },
+    {
+      "path": "checks-final.json",
+      "sha256": "sha256:7e61da6d5ba65fb2536ff97cd9ae39b450a1ccdf13bd0509cf62257d7cded4f6",
+      "bytes": 8212
+    },
+    {
+      "path": "checks.json",
+      "sha256": "sha256:d35f17c5a02d68ae611dd1a31bbe0fc58d558137055009f50d92b1ab54bd1faa",
+      "bytes": 4844
+    },
+    {
+      "path": "client-attempts/README.md",
+      "sha256": "sha256:043f25bbc8f39f1635690417cec17b364aae143f9b7e7eed22ea9c6e3cfeaa63",
+      "bytes": 1816
+    },
+    {
+      "path": "client-attempts/attempt-4.json",
+      "sha256": "sha256:a3c239c351fc74186f6505c6f5191427a2eee5f5ffecd3aabd3094950bbe695e",
+      "bytes": 68290
+    },
+    {
+      "path": "client-attempts/pilot-1.json",
+      "sha256": "sha256:ac60a729e355413d2a994a0922710b8a05446cb4180ec82a40652607c4442307",
+      "bytes": 2136
+    },
+    {
+      "path": "client-attempts/pilot-2.json",
+      "sha256": "sha256:1da7db1374ba158e51889894d6d46ab9fa1806ceaf1afe26a77841919e1852c4",
+      "bytes": 2384
+    },
+    {
+      "path": "client-attempts/pilot-3.json",
+      "sha256": "sha256:98137af994ea3f1f13b149142fee6c87042d378b3b3c91517d7bb4e7481e9ce8",
+      "bytes": 3835
+    },
+    {
+      "path": "current-calibration/baseline.json",
+      "sha256": "sha256:b033918d66d49b14596346b9c0b74559b2ca0f9071890acfb6ea893e7bced4bd",
+      "bytes": 917039
+    },
+    {
+      "path": "current-calibration/measurement.json",
+      "sha256": "sha256:10864787514bbaacdffb9404d8e8b1c7ceb7d1eb6373c92a4ddce6225dc76bba",
+      "bytes": 1142172
+    },
+    {
+      "path": "current-calibration/summary.json",
+      "sha256": "sha256:45fd4e0939e26e7ec256cb6a8b8352114e2ba414f2e2b4667385290ef16436dd",
+      "bytes": 3382
+    },
+    {
+      "path": "current-client-projection.json",
+      "sha256": "sha256:c003682047e9438d3cd2d4734164dd735d32d966288fc0d5cf087069266d4242",
+      "bytes": 80406
+    },
+    {
+      "path": "follow-up.md",
+      "sha256": "sha256:e5b72c41147fc848da5883bf4eaa661e138178c4456c522e38f61b1c401bd311",
+      "bytes": 6234
+    },
+    {
+      "path": "live-approval/README.md",
+      "sha256": "sha256:29c7e04c46b8e2ba66bfca93eaf887f75db96b0f0e303f3563e5bca74a51e45b",
+      "bytes": 2423
+    },
+    {
+      "path": "live-approval/budget.json",
+      "sha256": "sha256:a765198f0a7cefa2ae210166eb92a07e4edad2139023fae75c6bdeb629ae04dd",
+      "bytes": 290
+    },
+    {
+      "path": "live-approval/decision.json",
+      "sha256": "sha256:20364a42a7246c0101f7c60efe7d3b551f17fde2adbdb9144941a6d2d326e7da",
+      "bytes": 1602
+    },
+    {
+      "path": "live-approval/manifest.json",
+      "sha256": "sha256:71d088d4a5f2a102b1e61e09faecd7db34b5f47f95499c216661cea2d4f63012",
+      "bytes": 8473
+    },
+    {
+      "path": "live-package.json",
+      "sha256": "sha256:09c994902460839ab3d4376089fee7155c60b37353938523cc40f829b993b0ec",
+      "bytes": 25323
+    },
+    {
+      "path": "live-proposal.json",
+      "sha256": "sha256:95b1cf142ffeb1c343d6c0f9075597e1ce922222a78e4e58555d96e5b5ece50d",
+      "bytes": 12830
+    },
+    {
+      "path": "measurement.json",
+      "sha256": "sha256:41fe6c559448d30bf243c1b4221852267c2fe36dec91886f05b3e17fdc737b14",
+      "bytes": 979475
+    },
+    {
+      "path": "prior-calibration/README.md",
+      "sha256": "sha256:e485f991ba72ef509f10d509ee136624a48466e3e7c3924d3ce7dcc90fb8aa00",
+      "bytes": 1251
+    },
+    {
+      "path": "prior-calibration/integrity.json",
+      "sha256": "sha256:c7359db7a15ad1ae55e629ec66d6357abc08c2b781f485bc3a80188e1a3fb7fe",
+      "bytes": 1117
+    },
+    {
+      "path": "summary.json",
+      "sha256": "sha256:f97078610f94d702390f5a13cc4187d20e2c61475cbecc5c0701dcd243b042b6",
+      "bytes": 3371
+    }
+  ]
+}

package/docs/evaluation/token-footprint-2026-10-17/live-approval/README.md

@@ -0,0 +1,36 @@
+# Exact M1 pilot proposal — not approved
+
+The generated [manifest](manifest.json), [budget](budget.json) and
+[decision](decision.json) bind current source
+`sha256:33c450601727e38cf63b1874df0fcadc158c3d06af3b472b3d1603da6383c929`.
+The [full package](../live-package.json) includes metadata inventory, fixture/task
+hashes, routes and exact build/skill/matrix bindings. No real provider path ran.
+
+| Fresh Codex controller | Worker route | Starts | Work turns | Independent acceptance |
+| --- | --- | ---: | ---: | --- |
+| L1 | `review`: openai-codex / gpt-6.1-sol / high, read-only | 1 | 1 | T1 source claims and unchanged fixture |
+| L2 | `code`: openai-codex / gpt-6.1-sol / medium, coding | 1 | 1 | T2 fail before, pass after, only contract.mjs changed |
+| L3 | `review`, two sessions sharing immutable disposable T1 | 2 | 3 | Both reads, one follow-up in the first session, re-fetch original turn after follow-up and verify invariance |
+
+Controller: Codex 0.162.0 / openai / gpt-6.1-sol / high, with the hash-bound
+repository orchestration skill. Evaluation workers select no skills, extensions
+or external MCP grants. Production profiles and registrations are untouched.
+One pilot, three sequential fresh controllers, four starts/five work turns,
+zero whole-run retries. Turn supervision is 180 seconds after acceptance,
+case/controller 600 seconds, wave 1,800 seconds. Poll pauses are advisory;
+actual practice is measured. Cleanup uses separately bounded waits totaling
+60,500 configured milliseconds, plus process-table observation; no aggregate
+wall-clock hard cap or arbitrary subprocess containment is claimed.
+
+All approval flags are false; the advisory amount is null and unmetered use is
+not acknowledged. The existing gate requires an explicit current decision for
+this exact matrix/routes/budget, a positive **operator-selected advisory** USD
+amount and acknowledgment that subscription usage is unmetered. No reliable
+per-run cost or provider hard cap is available. An advisory is neither a tariff
+estimate nor a cap. Historical approval files are not reusable.
+
+After that decision, approve copies and recompute hashes of normalized validated
+manifest/budget objects, then bind the decision. No checked-in proposal can pass
+the live gate as it stands. Additional repetitions, A/B and Claude comparison
+require separate current approval. This package does not authorize publication
+or a production change.

package/docs/evaluation/token-footprint-2026-10-17/live-approval/budget.json

@@ -0,0 +1,13 @@
+{
+  "schema_version": "1.0.0",
+  "approved": false,
+  "manifest_digest": null,
+  "max_starts": 4,
+  "max_retries": 0,
+  "parallelism": 1,
+  "run_timeout_ms": 600000,
+  "total_timeout_ms": 1800000,
+  "advisory_usd": null,
+  "provider_cap_usd": null,
+  "acknowledges_unmetered_cost": false
+}

package/docs/evaluation/token-footprint-2026-10-17/live-approval/decision.json

@@ -0,0 +1,69 @@
+{
+  "schema_version": "token-live-decision-1",
+  "approved": false,
+  "manifest_digest": null,
+  "budget_digest": null,
+  "source_digest": "sha256:33c450601727e38cf63b1874df0fcadc158c3d06af3b472b3d1603da6383c929",
+  "build_digest": "sha256:f9d584a358e81ebfa36cf7e5d9f0ab212650f61217f375ddf1deefc7ff2190c1",
+  "matrix_digest": "sha256:9cdd89195dcb72d557c0b298bea796bdb8f2cfebb7ef8df249721596e572b27d",
+  "controller_skill_digest": "sha256:50878a59971d685021561c3d2d7757d1cbcfed9b2f5095202f9edfefe039837c",
+  "controller": {
+    "client": "codex",
+    "version": "0.162.0",
+    "provider": "openai",
+    "model": "gpt-6.1-sol",
+    "effort": "high",
+    "auth": "existing-client-login-not-probed"
+  },
+  "schedule": [
+    {
+      "id": "L1",
+      "task": "T1",
+      "starts": 1,
+      "turns": 1,
+      "route": "review",
+      "slots": [
+        "m1-l1"
+      ],
+      "prompts": 1
+    },
+    {
+      "id": "L2",
+      "task": "T2",
+      "starts": 1,
+      "turns": 1,
+      "route": "code",
+      "slots": [
+        "m1-l2"
+      ],
+      "prompts": 1
+    },
+    {
+      "id": "L3",
+      "task": "T1",
+      "starts": 2,
+      "turns": 3,
+      "route": "review",
+      "slots": [
+        "m1-l3-shared"
+      ],
+      "prompts": 3
+    }
+  ],
+  "budget_limits": {
+    "max_starts": 4,
+    "max_retries": 0,
+    "parallelism": 1,
+    "run_timeout_ms": 600000,
+    "total_timeout_ms": 1800000,
+    "provider_cap_usd": null
+  },
+  "advisory_usd": null,
+  "limits": {
+    "starts": 4,
+    "work_turns": 5,
+    "retries": 0,
+    "invocations": 3
+  },
+  "operator_acknowledges_advisory_not_cap": false
+}

package/docs/evaluation/token-footprint-2026-10-17/live-approval/manifest.json

@@ -0,0 +1,271 @@
+{
+  "schema_version": "1.0.0",
+  "approved": false,
+  "server_commit": "d6f360926e26629f7b382583c077016daeb954ca",
+  "screening_evidence": [],
+  "source_digest": "sha256:33c450601727e38cf63b1874df0fcadc158c3d06af3b472b3d1603da6383c929",
+  "tasks": [
+    {
+      "schema_version": "1.0.0",
+      "id": "T1",
+      "title": "Codebase explanation",
+      "family": "code-understanding",
+      "profiles": [
+        "P1"
+      ],
+      "views": [
+        "scout",
+        "docs-writer"
+      ],
+      "allowed_paths": [],
+      "prompt_hash": "sha256:d74a18bad9847c37741322c366c20ea05e6f2cc1bc046b164ae90ff2600c6000",
+      "fixture_hash": "sha256:579d5064e6529e3a08cbd7f3ecf9636f38ee138d17ee9b89d25fad81b8b69651",
+      "oracle_kind": "source-and-receipt",
+      "task_hash": "sha256:2b8a7d3cfbe985f25a9fe7e2c2e329e90f7c38a16f35ae3e180f1d9cd419531f"
+    },
+    {
+      "schema_version": "1.0.0",
+      "id": "T2",
+      "title": "Backend contract fix",
+      "family": "backend-fix",
+      "profiles": [
+        "P2",
+        "P3"
+      ],
+      "views": [
+        "implementer",
+        "reviewer"
+      ],
+      "allowed_paths": [
+        "contract.mjs"
+      ],
+      "prompt_hash": "sha256:3fb9a806d2c7de11f73c302088c5a3a1c48564f7588a7e392b6c98521840eccc",
+      "fixture_hash": "sha256:a5d8e952b1182e89398fff4fe0071d8537363aede0119f077ff4109f957a1b71",
+      "oracle_kind": "external-test-and-review",
+      "task_hash": "sha256:2a96d93af9503e98a474cecb240ecde2ee3052eaabef010fc16140a5b7123e7f"
+    },
+    {
+      "schema_version": "1.0.0",
+      "id": "T3",
+      "title": "Responsive accessible UI",
+      "family": "ui",
+      "profiles": [
+        "P4"
+      ],
+      "views": [
+        "ui-writer"
+      ],
+      "allowed_paths": [
+        "src/App.jsx",
+        "src/style.css"
+      ],
+      "prompt_hash": "sha256:cc72e58b0a7efcaf70170d339ac1197cd7e53fa862053107c6b9f4482323238c",
+      "fixture_hash": "sha256:fdce6e81a082a9e5cda3ad59c0c302e8a8919f2a20d5d6294c410583c7a699ae",
+      "oracle_kind": "browser-execution",
+      "task_hash": "sha256:fc513dbba4a2f8f90cca4494784d1a4832899b661f05d231c8db8d1fbeea562e"
+    },
+    {
+      "schema_version": "1.0.0",
+      "id": "T4",
+      "title": "Multi-module debugging",
+      "family": "debugging",
+      "profiles": [
+        "P2",
+        "P3"
+      ],
+      "views": [
+        "implementer",
+        "reviewer"
+      ],
+      "allowed_paths": [
+        "src/tax.mjs"
+      ],
+      "prompt_hash": "sha256:374611acced668d2e13e0bcd15eaf38749cb11888ea0791b4c1cd0878a1ab989",
+      "fixture_hash": "sha256:d98cb982a79a213408d181d1d30ba0051150c5762cf457f1cf9e01e5d14a037e",
+      "oracle_kind": "external-test-and-review",
+      "task_hash": "sha256:6b1a576b46aafea960738914848e68fb0499201f1345e3796579ab6a22a3e989"
+    },
+    {
+      "schema_version": "1.0.0",
+      "id": "T5",
+      "title": "Independent security review",
+      "family": "review",
+      "profiles": [
+        "P3"
+      ],
+      "views": [
+        "reviewer"
+      ],
+      "allowed_paths": [],
+      "prompt_hash": "sha256:97f8528b749b92838e0b3b4f4046a8fb81ffd9d14a60ecaeb33cf84c00a6858b",
+      "fixture_hash": "sha256:135a849ce7f8ceaec5a053466046b5e5d4f0d51227abca1038827e5ebb72a0cb",
+      "oracle_kind": "structured-findings",
+      "task_hash": "sha256:a76b843866bd20b4d40348575c7645e786947cc3eb49b0765b77f854c677e55e"
+    },
+    {
+      "schema_version": "1.0.0",
+      "id": "T6",
+      "title": "Regression test gap",
+      "family": "testing",
+      "profiles": [
+        "P4"
+      ],
+      "views": [
+        "tester"
+      ],
+      "allowed_paths": [],
+      "prompt_hash": "sha256:1d91df807efc91147ee7750d5730447d212de4ac9764f9a625a035e34fe82d68",
+      "fixture_hash": "sha256:e210625f68ee33354452e4ffa38d7a02759d7f1e480b6b60df9d668a5c3520b6",
+      "oracle_kind": "test-execution-and-coverage",
+      "task_hash": "sha256:8f15c34e5a1deffc17e754988e29aac62f1e28727985733019a728135b877433"
+    },
+    {
+      "schema_version": "1.0.0",
+      "id": "T7",
+      "title": "Infrastructure source review",
+      "family": "infrastructure-review",
+      "profiles": [
+        "P1"
+      ],
+      "views": [
+        "scout"
+      ],
+      "allowed_paths": [],
+      "prompt_hash": "sha256:cdacd6795b07e211e3a002a51aa49b2b1272e55cd5aec2aecc2b7e40851ef668",
+      "fixture_hash": "sha256:4053bd81269971ba8bae8b5bbc0066b43f50aaa8fcd5f638240fdd8b143a20f5",
+      "oracle_kind": "source-and-receipt",
+      "task_hash": "sha256:ef2252fc44f6750fd8630a0a48242b6ec602e5bfa7559226da6ec96aaf650fb3"
+    },
+    {
+      "schema_version": "1.0.0",
+      "id": "T8",
+      "title": "Evidence-based documentation draft",
+      "family": "documentation",
+      "profiles": [
+        "P1"
+      ],
+      "views": [
+        "docs-writer"
+      ],
+      "allowed_paths": [],
+      "prompt_hash": "sha256:2c35ae92d2aa1b62fdfe3c9bea0adb079d67ee9811db7c9a90971f0a84a87d0c",
+      "fixture_hash": "sha256:99aee6df8359ca777657bdc9fe6e5a09795e2b2a4d761c4919c5cb59ca950047",
+      "oracle_kind": "source-and-receipt",
+      "task_hash": "sha256:908d1af1652d71d8e5b917e16f656c62d443f97bae6000c7324c5a88763384a1"
+    },
+    {
+      "schema_version": "1.0.0",
+      "id": "T9",
+      "title": "Observed compaction continuation",
+      "family": "context-continuation",
+      "profiles": [
+        "P1",
+        "P2"
+      ],
+      "views": [
+        "scout",
+        "implementer"
+      ],
+      "allowed_paths": [],
+      "prompt_hash": "sha256:d8b100be1027e5eb54bd15581c3748fed50371dd5151165061cb53d9c0c4c709",
+      "fixture_hash": "sha256:c4d273343f9a6d324edfbf434fdacae1eb1bfc96a07cc08d992cf5ccead6aacb",
+      "oracle_kind": "observed-compaction-event",
+      "task_hash": "sha256:2567411f86cedee63fb7c601f786514351959ae2ac8e1485ef58775458e0d1b9"
+    }
+  ],
+  "runs": [
+    {
+      "run_id": "m1-l1",
+      "task_id": "T1",
+      "task_hash": "sha256:2b8a7d3cfbe985f25a9fe7e2c2e329e90f7c38a16f35ae3e180f1d9cd419531f",
+      "profile_id": "P1",
+      "view": "scout",
+      "repetition": 1,
+      "stage": "screening",
+      "route": {
+        "provider": "openai-codex",
+        "model": "gpt-6.1-sol",
+        "effort": "high",
+        "auth_type": "oauth",
+        "billing_type": "subscription",
+        "data_path": "synthetic-fixture-to-selected-provider"
+      },
+      "permission_profile": "read-only",
+      "variant": "A",
+      "review_target": null,
+      "skills": [],
+      "extensions": []
+    },
+    {
+      "run_id": "m1-l2",
+      "task_id": "T2",
+      "task_hash": "sha256:2a96d93af9503e98a474cecb240ecde2ee3052eaabef010fc16140a5b7123e7f",
+      "profile_id": "P2",
+      "view": "implementer",
+      "repetition": 1,
+      "stage": "screening",
+      "route": {
+        "provider": "openai-codex",
+        "model": "gpt-6.1-sol",
+        "effort": "medium",
+        "auth_type": "oauth",
+        "billing_type": "subscription",
+        "data_path": "synthetic-fixture-to-selected-provider"
+      },
+      "permission_profile": "coding",
+      "variant": "A",
+      "review_target": null,
+      "skills": [],
+      "extensions": []
+    },
+    {
+      "run_id": "m1-l3-a",
+      "task_id": "T1",
+      "task_hash": "sha256:2b8a7d3cfbe985f25a9fe7e2c2e329e90f7c38a16f35ae3e180f1d9cd419531f",
+      "profile_id": "P1",
+      "view": "scout",
+      "repetition": 1,
+      "stage": "screening",
+      "route": {
+        "provider": "openai-codex",
+        "model": "gpt-6.1-sol",
+        "effort": "high",
+        "auth_type": "oauth",
+        "billing_type": "subscription",
+        "data_path": "synthetic-fixture-to-selected-provider"
+      },
+      "permission_profile": "read-only",
+      "variant": "A",
+      "review_target": null,
+      "skills": [],
+      "extensions": []
+    },
+    {
+      "run_id": "m1-l3-b",
+      "task_id": "T1",
+      "task_hash": "sha256:2b8a7d3cfbe985f25a9fe7e2c2e329e90f7c38a16f35ae3e180f1d9cd419531f",
+      "profile_id": "P1",
+      "view": "scout",
+      "repetition": 2,
+      "stage": "screening",
+      "route": {
+        "provider": "openai-codex",
+        "model": "gpt-6.1-sol",
+        "effort": "high",
+        "auth_type": "oauth",
+        "billing_type": "subscription",
+        "data_path": "synthetic-fixture-to-selected-provider"
+      },
+      "permission_profile": "read-only",
+      "variant": "A",
+      "review_target": null,
+      "skills": [],
+      "extensions": []
+    }
+  ],
+  "prerequisites": {
+    "baseline_passed": false,
+    "screening_passed": false,
+    "bottleneck": null
+  }
+}