R6 (finish-operator 2026-08-02): the only envelope number anyone quoted
was a stale fixture comment (workflows/stopship.workflow.js:1-5,
17,457–17,550 tokens from an old live run). Nothing failed when the
assembled payloads grew. Three regressions now build — never send —
the real payloads and pin them at measured-current +10%:
- read_only_child_envelope_stays_within_measured_ceiling: scout system
prompt + seed messages + child tool catalog = 80,856 bytes measured
(~20k tokens — the old 17.5k figure was already low); ceiling 89,000.
- parent_agent_surface_stays_within_measured_ceiling: BASE_PROMPT +
full parent tool catalog = 72,679 bytes measured; ceiling 80,000.
- turn_meta_block_stays_within_measured_ceiling: per-turn <turn_meta>
block via the production snapshot path = 254 bytes measured
(includes the new sandbox-posture line); ceiling 280.
Each failure message says how to raise the ceiling honestly (re-measure
in the same commit). The stopship fixture comment now points at the
pinned test instead of the stale token figure.
Proof the ceilings bite: a scratch diff padding BASE_PROMPT +15,000B,
the child system prompt +17,000B, and turn_meta +1 line (~+20% each)
failed all three tests; reverted, all green.
Verification:
- cargo fmt --check -p codewhale-tui: clean
- cargo test -p codewhale-tui --bin codewhale-tui -- turn_meta_block_stays
read_only_child_envelope parent_agent_surface: 4 passed; 0 failed
Land the in-flight experimental-search slice as a real, honest v0.9.4
piece: a provider-neutral WorkflowSearchSpec authoring + freeze boundary
(parse/validate, preregistration hashes over baseline/requested+resolved
model/public evidence/evaluator, deterministic candidate ids and
admission batches) plus the operate best-of-N recipe's structured
'search' strategy (2-16 independent worktree candidates with
responseSchema contracts and a read-only judge). The module explicitly
remains an authoring boundary, not a runtime: hard_gates/score commands
are parsed and validated only, and docs say so.
Fixes found during the takeover operation:
- The checked-in recipe test failed against the real driver contract
(parse_task_options lets prompt win over description, so fake-driver
needles never matched and replies fell back to non-JSON). The recipe
now puts the full instruction in the single driver-visible description
and pins the prompt-wins contract with a new test.
- Queue claims verified against the runtime: the Workflow host's
per-run concurrency gate (Semaphore, 16 live) is where larger
populations wait; docs now name the gate instead of vague 'queues
through Fleet'. WORKFLOW_SEARCH_MAX_CONCURRENT documents 16 as today's
default with a cross-reference (a crate cycle prevents importing the
host constant).
- TournamentOrdering (ScoreThenCost) and the 1,000-agent validation
test bump retained from the slice.
Verified: cargo test -p codewhale-workflow -p codewhale-workflow-js
--locked green (250 + 16 + 9 + 49).
(cherry picked from commit f3e3232ef2)
Signed-off-by: Hmbown <101357273+Hmbown@users.noreply.github.com>
Remove the remaining RLM and Work compatibility registrations, route bang-shell execution through Bash, and migrate engine, Fleet, workflow, and PTY fixtures to canonical action calls. Preserve apply_patch for the DeepSeek Responses custom-tool contract and teach repository-law enforcement to classify canonical File mutations.
* chore(workflows): delete the seven v0.8.68 lane scripts
v0868_catalog_lane, v0868_issue_implement, v0868_issue_sweep,
v0868_release_gate, v0868_stopship_lane, v0868_tui_copy_lane, and
v0868_workflow_ui_lane are pinned to v0.8.68 issue numbers that are closed.
They were the first-generation workflow scripts, written while we were still
learning the shape from grok-build. The generic replacements they taught us to
write already ship: operate_best_of_n, operate_parallel_scout,
operate_read_audit, operate_staged_fix, stopship, issue_audit.
Per docs/AGENT_ETHOS.md a workflow script is a diff, not intent. The intent of
each lane is preserved in its closed issues; keeping version-pinned scripts is
the same habit that produced 445 branches.
Six had no inbound references. The seventh did:
short_workflow_names_do_not_resolve_historical_v0868_files required
workflows/v0868_issue_sweep.workflow.js to exist on disk, so deleting it alone
would have failed cargo test -p codewhale-cli. That test covers two behaviors
and only one was v0868-specific, so it keeps the guard that a bare short name
never expands to a version-pinned script, and checks explicit-path resolution
against stopship.workflow.js instead. Renamed to match what it now asserts.
cargo test -p codewhale-cli: 169 passed, 0 failed.
* style(cli): cargo fmt the workflow path resolution test
type "general" maps to a write-capable worker and fails closed without
writeRoots/exactFiles/coordinationContracts. operate_read_audit and
operate_parallel_scout only synthesize scout output — use type "review"
so the starters complete under Operate dogfood.
Rewrite OPERATE_MODE into hard conductor doctrine (goal → dispatch →
background → verify → synthesize). Dispatching sub-agents is the default
multitask path — no /multitask verb. Write-capable children get a spawn-time
VERDICT PASS/FAIL evidence contract; the agent tool description prefers
implementer/verifier pairs and worktrees. Ship operate_* starter workflows
and strengthen the best-of-n skill so winners apply only after PASS.
Keep every acceptance role on the same five-file source boundary and carry all seven runtime owners through each promoted handoff. Pair the Workflow completion constructor with its completed-path assertion so a live reviewer can distinguish source emission from terminal proof.
Add the tmux process-exit reconciliation owner to the bounded scout query and require every downstream receipt to distinguish Workflow completion from Lane reconciliation. Lock the five-file, seven-owner contract into fixture tests.
Propagate a terminal blocking gate into the workflow run result instead of accepting a successful VM return.
Add the release-lead final receipt to the stopship fixture, constrain scout evidence to the exact four source files, and reject placeholder-only gate artifacts with focused regressions.
Tell the bounded release scout to follow its six-owner decision rule literally and reject only artifacts that name an actual missing owner. Lock the contract into the JavaScript authoring fixture test.
Keep the scout focused on runtime source owners while the already-typed run_started event remains authoritative for the compiled workflow id and source path.
Use one exact-symbol, bounded repository grep for stopship source evidence so model providers cannot serialize per-file searches into an exhausted response budget.
Fail closed when an explicit approval would promote an empty declared artifact, and require every tool-free stopship consumer to carry concrete source evidence into its next handoff.
Add an explicit deny-all tool posture for declarative leaves and use it after the scout so stopship roles review promoted evidence instead of reopening source discovery until their step caps.
Give each bounded stopship role a small response reserve for the required terminal verdict while keeping all token and wall-time ceilings unchanged at a 360k aggregate cap.
Make every stopship role begin its final response with the host-readable verdict token and lock the anti-preamble instruction into the checked-in fixture contract.
Constrain each live acceptance role to exact-file retrieval rounds and a final no-tool verdict within its authored model-turn cap. Keep the existing fail-closed token budget while preventing broad search fan-out from consuming every response before APPROVE or BLOCK.
Integrate the underwater TUI, message-first Operate, Fleet and Workflow reliability, expanded model/provider catalog, exact custom-route restoration, docs-first site, localization, packaging, and release metadata for the v0.9.0 candidate.
Harden endpoint-bound credential provenance, approval and goal UX, Fleet attempt fencing and crash recovery, large-workspace mention discovery, Kimi budgeting, and release asset/version gates. Include the stopship Fleet and Workflow fixtures used by release dogfood.
Verified with workspace fmt/check/clippy/tests on Rust 1.88, release-script and npm suites, 18-crate publish dry run, production web build, Docker build check, secret scan, dependency audit, and protected-state hash validation.
Base each role cap on the 17.5k-token live first-turn envelope plus bounded per-turn headroom. Reduce the step ceilings, document the 360k aggregate/shared limit, and lock the contract into the JavaScript authoring regression.
Add an opt-in gate contract that requires a standalone first-line verdict and blocks missing or malformed responses without changing legacy pass-on-success gates. Apply it to every v0.8.68 acceptance handoff.
Keep role-only Fleet leaves type-free while preserving authored non-General types so compatible overrides remain visible and conflicts still fail before provider dispatch.
Tests: cargo test -p codewhale-workflow --lib --locked
Tests: focused codewhale-tui verdict, role-only Fleet, and stopship fixture tests
Make every acceptance role return an exact first-line APPROVE or BLOCK verdict, direct source-reading children to grep before bounded snippets, and replace the unverified budget inflation with explicit per-role caps. Structural and mock receipt tests now enforce the prompt and budget contract; live provider completion remains unverified.\n\nTests: cargo test -p codewhale-workflow --lib --locked stopship_acceptance_fixture_is_read_only_and_gate_complete -- --nocapture\nTests: cargo test -p codewhale-tui --bin codewhale-tui --locked stopship_acceptance_fixture_emits_role_gate_and_terminal_receipts -- --nocapture
The first real tmux run reached the scout Fleet role but exhausted the 6k child budget before completing its source inspection. Raise the bounded per-role token caps enough to cover the runtime prompt and read-only evidence pass, and lock the minimum into the fixture contract test.
Replace the stale closed-issue fix workflow with a five-role, no-edit acceptance chain and explicit Workflow-owned gates. Allow a Fleet implementer profile to be narrowed to read-only tools, keep custom Fleet roles compatible with general profiles, and remove the unroled reducer that made named-Fleet runs fail after all children completed.\n\nAdd fixture-level contract coverage plus an exact mock runtime test for role/profile, gate, and terminal receipts, and refresh the v0.8.68 playbook with the disposable-workspace command.
Align checked-in v0.8.68 workflow fixtures with the typed TaskMode schema. Keep the dogfood verifier read-only so Fleet validation can reach runtime execution.
Add a workflow run CLI entrypoint that validates a checked-in workflow and named fleet, then starts a Runtime-backed Lane using the existing headless Workflow tool path. Wire named fleet role maps into the TUI Workflow driver so task({ role }) resolves to the fleet-selected profile before sub-agent spawn. Update the stopship dogfood workflow to declare roles and refresh the playbook target command.
Add fleets/v0868-stopship.toml (five roles), NamedFleet loader with
resolution tests, profile bindings on stopship workflow steps, and
playbook commands for lane + interim exec paths.
Add wave-based Workflow JS files for stopship, catalog, workflow UI,
TUI copy, release gate, and per-issue implementation. Include agent
playbook, milestone label sync action, and update agent-task template.
Document main as v0.8.68 source of truth and defer v0.8.69 refactors
until stopship (#4090, #4093, #4094) is green.
Co-authored-by: CodeWhale Agent <codewhale-agent@hmbown.local>
Safety checkpoint of multi-agent cutover work (fleet roster core,
whaleflow-js runtime, /fleet roster view, audit fix lanes). Tree may
not compile: the agent-tool profile edit in tools/subagent/mod.rs was
interrupted mid-edit (credit exhaustion). Checkpoint precedes repair.
Adds a compile-only JavaScript/TypeScript authoring path that extracts a JSON-compatible workflow({...}) object, lowers it to the existing WorkflowSpec IR, and runs the Rust validation gate before execution.
Includes a branch/reduce .workflow.js example and a short authoring design note comparing YAML/JSON, Starlark, JavaScript, and TypeScript. The compiler rejects effectful JavaScript constructs instead of executing workflow source as a second runtime.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add a fail-closed Starlark authoring layer that compiles workflow files into WorkflowSpec without exposing runtime execution. Include ctx.* repair aliases, unsupported-construct rejection, and rlm_cache_change / issue_fix_tournament examples.
Refs #2670.
Preserves WhaleFlow direction credit for @AdityaVG13 in changelogs and PR notes.
Hmbown