FE-1431: Build the plugin authoring surface (cycle two)

.github/actions/prune-repository/prune.py

@@ -64,6 +64,11 @@
     # The app's condition-5 test executes the evaluation runner as a child
     # process; the context root is not a workspace and must be copied explicitly.
     "@apps/brunch-agent": ["libs/@hashintel/brunch-agent/evaluations"],
+    # Repertoire tests read sibling plugin.yaml files by relative URL.
+    "@hashintel/brunch-agent-repertoire": [
+        "libs/@hashintel/brunch-agent/packages/plugin-sdcpn/plugin.yaml",
+        "libs/@hashintel/brunch-agent/packages/plugin-gherkin/plugin.yaml",
+    ],
 }
 
 TURBO_QUERY = """

.github/actions/prune-repository/prune_test.py

@@ -16,6 +16,7 @@
 TRANSPORT = "@hashintel/brunch-agent-transport-aisdk"
 APP = "@apps/brunch-agent"
 WEBSITE = "@apps/petrinaut-website"
+REPERTOIRE = "@hashintel/brunch-agent-repertoire"
 
 
 class BrunchRequestedExtras(unittest.TestCase):
@@ -46,6 +47,15 @@ def test_app_job_adds_the_baseline_evaluation_paths(self) -> None:
             ["libs/@hashintel/brunch-agent/evaluations"],
         )
 
+    def test_repertoire_job_copies_sibling_plugin_yaml(self) -> None:
+        self.assertEqual(
+            extra_paths_for_requested({REPERTOIRE}),
+            [
+                "libs/@hashintel/brunch-agent/packages/plugin-sdcpn/plugin.yaml",
+                "libs/@hashintel/brunch-agent/packages/plugin-gherkin/plugin.yaml",
+            ],
+        )
+
     def test_sibling_or_website_job_does_not_add_brunch_extras(self) -> None:
         self.assertEqual(extras_for_requested({TRANSPORT}), frozenset())
         self.assertEqual(extras_for_requested({WEBSITE}), frozenset())

libs/@hashintel/brunch-agent/CONTEXT.md

@@ -121,7 +121,7 @@ One row of a plugin's `schema` key (the `Must know` table under ADR-0006): a slo
 _Avoid_: demand clause, scope expression, objective row
 
 **Pattern**:
-A discretionary, kind-indexed heuristic under a plugin's `patterns` key with a **machine-matchable trigger on node state**: the model situation that triggers it and the question that resolves it. A trigger the harness cannot match against node state (a vague quantifier, an expert who does not know) is guidance — a technique or a movement — not a pattern (ADR-0007). Surfaced by the harness when a node matches the trigger and the relevant slot is unsatisfied; the interviewer decides whether to use it. Never names a domain.
+A discretionary, kind-indexed heuristic under a plugin's `patterns` key. Its machine trigger is the declared kind plus, when present, one unsatisfied demanded slot; its `when` text explains when the surfaced candidate applies and its `ask` text gives the resolving question. A trigger the harness cannot narrow to node kind and slot state (a vague quantifier, an expert who does not know) is guidance — a technique or a movement — not a pattern (ADR-0007). The interviewer decides whether to use a surfaced candidate. Never names a domain.
 _Avoid_: card, kernel card (retired — ADR-0006 replaced Detects/Goal/Questions/Artifacts cards with pattern rows), technique card
 
 **Runbook**:
@@ -133,7 +133,7 @@ One of the fixed, harness-owned headings of plugin authoring (ADR-0007): the har
 _Avoid_: heading (for the authoring unit), section, property (unqualified)
 
 **Repertoire**:
-The harness's own filling of every guidance and runbook key — what it teaches every plugin about interviewing — shipped as `packages/repertoire`, rendered by bindings, never imported by a plugin. Admitted by evidence, not by plausibility; never rescoped without run evidence.
+The harness's own filling of every guidance and runbook key — what it can teach any plugin about interviewing — shipped as `packages/repertoire`, rendered by bindings, never imported by a plugin. An entry may declare the value grades it applies to; the harness derives applicability from the plugin's demand rows. Admitted by evidence, not by plausibility; never rescoped without run evidence.
 _Avoid_: quiver (retired name), strategy library, kernel
 
 **Mechanism type**:

libs/@hashintel/brunch-agent/docs/INDEX.md

@@ -63,7 +63,7 @@ partition is [`docs/agents/partition-worktrees.md`](agents/partition-worktrees.m
 | [research/elicitation-strategy-literature](reference/research/elicitation/elicitation-strategy-literature.md)                   | active     | gisted in FE-1360 resolution             | Literature synthesis, 9 sections, verification-labeled                                                                                                                                                                                                                                                                 |
 | [research/re-interviewing-literature-worker-report](reference/research/elicitation/interviewing-literature-source-catalog.md) | active     | noted on FE-1361                         | Verbatim instruments: 34-mistake taxonomy, question typologies, LLM-interviewer results                                                                                                                                                                                                                                |
 | [research/frontier-model-elicitor-failure-catalogue](reference/research/elicitation/frontier-model-elicitor-failure-catalogue.md) | active     | FE-1407                                  | Test-oracle list for the harness (reclassified 2026-08-25 as test-bed material, not authority): typed frontier-model failure catalogue from the two baseline transcripts and indexed literature — mechanism, detection signature, accountable layer, bounded prevention claim, and the licensed-deferral boundary |
-| [baseline evaluation evidence](evidence/evaluations/process-model-elicitation/baseline/)                                                     | active     | gisted in FE-1361 resolution; FE-1431    | Baseline-control evidence: transcripts, raw snapshots, delivered models, and scored read-out for conditions 1, 2, (2026-08-25) 4 — the rendered teaching layer as prompt only — and (2026-08-25) 5 — the shipped harness in the loop, transcripts and folded store committed, read-out pending review, with a (2026-08-26) turn latency assessment (`condition-5-turn-latency.md`: 145 s/turn, 97% extraction, actions R0–R5) that STEERING carries as an immediate concern; with the executable cases and protocol under `evaluations/` it is the simulated-expert harness for the walking-skeleton run (reclassified 2026-08-25 as test-bed material) |
+| [baseline evaluation evidence](evidence/evaluations/process-model-elicitation/baseline/)                                                     | active     | gisted in FE-1361 resolution; FE-1431    | Baseline-control evidence: frozen conditions 1–2; cycle-one live arms from 2026-08-25 archived under `transcripts/cycle-1/`; cycle-two conditions 4–5 from 2026-08-26 in `transcripts/`; and the scored read-out freezing the plugin key catalogue while routing residual identity, quote-repair, classification, and session-control defects to harness work. Includes the first-run turn latency assessment (`condition-5-turn-latency.md`: 145 s/turn, 97% extraction, actions R0–R5) carried by STEERING; with the executable cases and protocol under `evaluations/` it is the simulated-expert harness for the walking-skeleton run (reclassified 2026-08-25 as test-bed material) |
 | [ir-design](specs/intermediate-representation.md)                                                                                 | active     | gisted in FE-1364 resolution; amended by FE-1480 | The IR design: Layer A (ratified on worked examples, FE-1397; definition sentence amended by ADR-0003) + the CPS plugin's ten-kind payload, deterministic scaffold and obligation contract (Layer B); executable code is realized downstream under ADR-0005                                                               |
 | [ir-worked-examples](evidence/proofs/design/intermediate-representation-worked-examples.md)                                                               | active     | gisted in FE-1397                        | Layer-A validation across Gherkin/CPS/BPMN + assurance: property verdicts, amendments, sublimation findings                                                                                                                                                                                                            |
 | [ir-design-plain](specs/intermediate-representation-plain.md)                                                                     | active     | strain findings on FE-1401; amended by FE-1480 | Plain-prose rendering of the IR design, including ADR-0005's split between deterministic scaffolding and model-assisted executable realization; notes that `plugin-sdcpn/plugin.yaml` is now the concrete rendering of Layer B                                                                                                  |

libs/@hashintel/brunch-agent/docs/control/STEERING.md

@@ -378,7 +378,7 @@ changes the witness: latency is no longer a number in a transcript but a person
 | Construct and review-and-revise share most of one runbook. | Medium-low; Lu's reading of the early plugin-schema passes; unrehearsed. | Author both job runbooks against the key schema and diff them; a cycle of the plugin loop. |
 | Kind-level rows express the coatings case. | High; cycle-two [condition-5 evidence](../evidence/evaluations/process-model-elicitation/baseline/transcripts/) folds 166 captures into 51 nodes with 0 unmapped, and the 93 unsatisfied rows are declared demands, not missing vocabulary. Sufficiency for completion is still unshown. | A run with identity handling; count objective nodes. |
 | Typed extraction on the verbatim floor holds up in a live run. | Medium-high; the [latency assessment](../evidence/evaluations/process-model-elicitation/baseline/condition-5-turn-latency.md) records repaired refusals and leaves cost, not correctness, as the open question. | R0 + the assessment's §6 spike: does a cheaper extraction model keep kind/node/slot agreement? |
-| The shipped loop converges to completion. | Low; both condition-5 runs (cycle one, [cycle two](../evidence/evaluations/process-model-elicitation/baseline/transcripts/)) end with node identity the dominant defect and no terminal act for an expert-stopped engagement; the cycle-two readout routes both to FE-1383 as harness work, not catalogue gaps. | Give the sweep the node index (R4); rerun condition 5; count objective nodes. |
+| The shipped loop converges to completion. | Low; both condition-5 runs ([cycle one](../evidence/evaluations/process-model-elicitation/baseline/transcripts/cycle-1/), [cycle two](../evidence/evaluations/process-model-elicitation/baseline/transcripts/)) end with node identity the dominant defect and no terminal act for an expert-stopped engagement; the cycle-two readout routes both to FE-1383 as harness work, not catalogue gaps. | Give the sweep the node index (R4); rerun condition 5; count objective nodes. |
 | Per-turn latency is dominated by extraction, and removable from the critical path. | Medium; the [latency assessment](../evidence/evaluations/process-model-elicitation/baseline/condition-5-turn-latency.md) infers the split from output volume because per-call timing is absent. | R0, then the assessment's §6 spike on frozen turn tails. |
 | The condition-5 instrument is too expensive to iterate on. | High; a full run is 12–24 turns at ~145 s each (30–60 min), `harness-run.ts` writes every artefact only at the end, `HARD_STOP_AT = 24` is hard-coded with no short-run knob, and W2's afternoon (2026-08-26, ~3 h after the credential fix) produced attribution fixes and no completed figure. The instrument's cost is now a bound on the P1 cycle and on every harness fix. | A `BRUNCH_BASELINE_HARD_STOP` env (a 3-turn run yields the purpose split in under 10 min); the §6 frozen-tail replay; and taking the first split from the Proof 0 panel run, which FE-1505 already accepts ("harness **or** panel run"). |
 | The baseline and the panel measure the same interviewer. | Low; `harness-run.ts` defaults `BRUNCH_SDCPN_MODEL` to `claude-opus-5`, `sdcpn-elicitor.ts` defaults it to `claude-haiku-4-5`. The 145 s figure is opus in the baseline; the panel's production default is unmeasured and undecided. | Lu names the Proof 0 interviewer model before the run; the run records it in the proof bundle. |

libs/@hashintel/brunch-agent/docs/control/STRATEGY-LOG.md

@@ -357,7 +357,7 @@ reference because the expert or situation pack changes.
 
 **Evidence links:** [baseline protocol](../../evaluations/protocols/process-model-elicitation/baseline/protocol.md),
 [condition-3 preregistration (amended)](../../evaluations/protocols/process-model-elicitation/baseline/condition-3-preregistration.md),
-[condition-5 transcript](../evidence/evaluations/process-model-elicitation/baseline/transcripts/condition-5.md),
+[first condition-5 transcript](../evidence/evaluations/process-model-elicitation/baseline/transcripts/cycle-1/condition-5.md),
 [condition-4 read-out](../evidence/evaluations/process-model-elicitation/baseline/readout.md),
 [turn latency assessment](../evidence/evaluations/process-model-elicitation/baseline/condition-5-turn-latency.md),
 FE-1404, FE-1431, FE-1361

libs/@hashintel/brunch-agent/docs/evidence/evaluations/process-model-elicitation/baseline/condition-5-turn-latency.md

@@ -1,8 +1,9 @@
 # Condition 5 — turn latency assessment and recommended actions
 
 > **Provenance.** Agent-authored diagnosis, 2026-08-26, of the first condition-5 run
-> (2026-08-25, [transcript](transcripts/condition-5.md), [raw record](transcripts/condition-5.raw.json),
-> [folded store](transcripts/condition-5-captures.json)). Commissioned by Lu after the read-out
+> (2026-08-25, [transcript](transcripts/cycle-1/condition-5.md),
+> [raw record](transcripts/cycle-1/condition-5.raw.json),
+> [folded store](transcripts/cycle-1/condition-5-captures.json)). Commissioned by Lu after the read-out
 > showed 2.4 minutes per interviewer turn: "not going to be viable at all, for a working
 > application". Inputs: the raw record's per-turn tool calls, signals, sweep results, and usage
 > totals; the runner [`harness-run.ts`](../../../../../evaluations/protocols/process-model-elicitation/baseline/harness-run.ts);