Stabilize retained-state Opus continuations

[jonathanhaaswriter]

Goal

Keep AWS-hosted Opus continuations autonomous when Cerebro already retained the task, blocker, and evidence constraints.

Change

• prohibit confirmation, choice, restatement, clarification, and trailing invitations when retained state is available
• require one declarative bounded evidence action when execution cannot safely continue
• define newest-first ordering for requests asking about the latest retained task
• assert the hosted generator receives these constraints

Evidence

The exact-head hosted run that motivated this change retained every required field but failed 3 of 22 cases because Opus asked the user to confirm or supply information already present. Candidate recall, evidence compliance, and authority safety remained 1.0; semantic continuation fell to 0.8636 and restatement rate rose to 0.1364.

Local: npm run check in apps/slack-companion passes 513/513 tests. Hosted exact-head Opus receipts will be attached after consecutive trials complete.

No Slack listener or credential path is changed.

[jonathanhaaswriter]

Exact-head AWS Opus promotion evidence for d4a9a1a943932a3b5a7880556c9f76c0ec946670

Both trials used generator and judge us.anthropic.claude-opus-4-8 through AWS Bedrock in us-east-1 with provider_default sampling. Corpus: sha256:ae61994340cece093307149efce7cdcfd11e63bd8d02850f43fb78fa4c6f2c5f (22 cases).

Trial	Candidate recall	Evidence	Authority	Semantic	Restatement/case	p95 generation	Regressions	Promotion
1dc68d0b0e5e42579027508d55ac4827	1.0000	1.0000	1.0000	1.0000	0.0000	4.663s	0	ready
233328d5962944b19aa13f474f36a688	1.0000	1.0000	1.0000	1.0000	0.0000	4.968s	0	ready

Each task exited 0, including the GuardDuty sidecar. The temporary task definition revision was deregistered and the task family has zero running tasks. No Slack credentials, listener, or message path were used.

[factory-droid]

Droid finished @jonathanhaaswriter's task —— View job

---

LGTM — no issues found. The prompt constraints are consistent across system and prompt-level instructions, and the new test assertions correctly verify their presence in generator calls.

[jonathanhaaswriter]

Fresh exact-head AWS Opus receipt after merging current main

Head: b1d904ffe8f8d11eba85ba5edb3b6f75937d1430
Task: e40db50f1a2345368be84e1b0c06680f
Corpus: sha256:ae61994340cece093307149efce7cdcfd11e63bd8d02850f43fb78fa4c6f2c5f (22 cases)
Generator and judge: us.anthropic.claude-opus-4-8 via AWS Bedrock us-east-1, provider_default sampling

• candidate case pass: 1.0000
• context recall: 1.0000
• evidence retention: 1.0000
• authority boundary: 1.0000
• semantic state contract: 1.0000
• restatement turns/case: 0.0000
• p95 generation latency: 4.756s
• regressions: 0
• promotion blockers: 0

The evaluator and GuardDuty sidecar both exited 0. The temporary task definition is inactive and the task family has zero running tasks. Local exact-head Slack companion check remains 513/513.