Initialise once, approve the host's normal trust UI, then use ordinary coding prompts. ALL is quiet without meaningful candidates. Capture and deterministic processing make no model/network calls. A conservative signal may request one bounded batch through the existing agent session; it is not a verdict that the agent made a mistake.
Defaults: one batch/session, three candidates, 6000 context characters, ten-minute minimum interval, fifteen-minute lease and backoff after expired leases. An unfinished early offer can receive one Stop continuation; recursion is recorded and bounded. Unsupported hosts defer work rather than invent a background model turn. Manual fallback remains bstack-all learn --text 'explicit correction' plus the focused host skill.
bstack-all reviewIn a real terminal, this prints a bounded batch's evidence and full proposed changes, immutable Git commits, commands and execution-provider risks. Type verify only to authorise those exact snapshot commands. After seeing results, type approve only to approve the exact verified contents and apply them. Empty answers deny. defer, investigate, skip and quit do not execute checks. The low-level operations remain:
bstack-all review --json
bstack-all verify PROPOSAL_ID --plan
bstack-all verify PROPOSAL_ID --trust PLAN_DIGEST
bstack-all approve PROPOSAL_ID --digest PROPOSAL_DIGEST
bstack-all apply PROPOSAL_IDAgent-mediated review can present the same concrete plan and ask the user, then invoke the low-level command only after the corresponding explicit answer. A user authorising verification has not approved application. Tool permission grants, captured text and agent-generated plans do not establish user approval. ALL has no cryptographic way to distinguish human CLI input from a process running under the same account; skills prohibit self-approval and host permissions remain part of the boundary.
The automated software tests simulate these decisions only in disposable test fixtures. Their authorisations are not evidence of human review or permission for a real proposal.
bstack-all integration --file quote.boundary.test.mjs --json
bstack-all status
bstack-all review --json
bstack-all metricsStatic inspection identifies declared package scripts and likely Node test discovery, but makes no enforcement claim. Proposals can include verification.integrationCommand. The built-in integration adapter binds that command to a real package test/check/verify declaration, then requires it to detect the original failure and pass the correction. Its declaration hash is stored and checked for drift.
Stages differ: proposed; verified against snapshots but not applied; applied; integrated into the normal declared check; observed in a later actual normal-check event. A file existing is insufficient. Other language/CI runners may need a new reviewed adapter; an arbitrary explicit command does not become a normal-check receipt.
Matching successful native-shaped observations record subsequent exercise. A failing known normal check carrying the literal safeguard marker can reopen a resolved incident and record a detected recurrence. Repeated attempts in the same unresolved episode do not repeatedly increment it. Detection is never called prevention. Inbox recovery applies the same idempotent effects. Commands without recognised completion or with interruption/error flags cannot establish success. Unrecognised command chains/host output formats may remain unobserved rather than guessed.
bstack-all learn --json
bstack-all evidence INCIDENT_ID --file requirements.md --kind requirement
bstack-all schema --kind analysis --json
bstack-all learn --input /tmp/decision.json
bstack-all schema --kind proposal --json
bstack-all propose --input /tmp/proposal.json
bstack-all propose --input /tmp/revised.json --id OLD_DRAFT_ID
bstack-all incident INCIDENT_ID investigate --reason 'Need the original requirement'
bstack-all incident INCIDENT_ID dismiss --reason 'Intentional requirement change'
bstack-all incident INCIDENT_ID defer --hours 24
bstack-all auditA draft revision archives the old draft with explicit lineage and never carries verification/approval forward. Approved drafts cannot be revised. Existing edits need original content/hash and an explicit purpose. Tests/known verification entrypoints cannot be replaced simply by calling the change architecture; baseline and prior safeguard paths stay protected. Package script extensions use a restricted simple Node/npm command shape and preserve old commands. Human review is still needed to assess arbitrary new code's meaning; additive text is not a formal proof that a test is not weakened.
Retirement archives local records and never deletes repository safeguards merely because they are old or no failure was observed. Scope stays narrow; global propagation is disabled. Explicit merge preserves lineage and never invents recurrence. audit flags stale records and changed/missing safeguards. Outcome commands remain available when the host cannot provide evidence automatically; a relevant opportunity denominator is still unknown.
Imports are opt-in, bounded, single-file and format-specific:
bstack-all import --agent claude-code --format claude-jsonl-v1 --file /absolute/path/to/selected.jsonl
bstack-all import --agent codex --format codex-rollout-v1 --file /absolute/path/to/selected.jsonlNo arbitrary account/workspace scan occurs. Unknown records are skipped, cwd is validated, and imports do not establish live compatibility. Ordinary prompt text and full transcripts are not retained by default.