Closes the second and final ask from the C4 cross-review (Karpathy
specific ask #1). The previous non-ceo-dm-reply recording (409bec2)
pre-dated C4 part 3/5 by ~18 minutes, so it could not discriminate
between "non-CEO path correctly does not receive hire_agent /
file_approval" and "the tools simply had not been injected yet."
Re-recorded against anthropic/claude-sonnet-4-5 on the current main.
The VP Product agent (manager — canRequestHire=true by keyword match,
but after F3 canFileCeoStrategy=false because reportsTo is not null
AND role is not CEO) produces:
- mark_read + /api/agent_message/send + m_77 preserved — 3 positive
assertions intact.
- Zero hire_agent invocation (not asked to hire).
- Zero file_approval invocation — positive evidence the F3 split
correctly excludes file_approval from non-CEO managers. Without
F3 this call would have had file_approval in scope and a manager
under pressure might have reached for it.
Fixture assertions updated to pin the two tool-scope invariants using
the new message_does_not_contain_substring kind from F2.
C4-followup.md marked CLOSED with resolution table mapping each of
the four items to its landing commit, plus the companion cleanup
(43337e5 — RpcMethod single-source-of-truth) that came out of the
same session.
Final: typecheck green, 537 unit tests, 10 eval tests.
Closes the fourth ask from the C4 cross-review (docs/proposals/C4-followup.md
Item 4). The eval harness previously had no way to assert the absence of a
pattern in a trace — only positive `message_contains_substring`. That meant
the C4 "no more node -e / bash+fetch from CEO" invariant was only
demonstrated by the diff between two commits, not by any running assertion
that future regressions would trip.
Three changes:
1. Extend the Assertion union in tests/eval/types.ts with
`message_does_not_contain_substring` (role + substring). Symmetric to
the positive kind; fails on first match with a specific reason.
2. Add the case in assertion-judge.ts. Same traversal as the positive
branch but inverted exit condition.
3. Add two harness self-tests covering pass + fail, so the new kind is
exercised even when the fixtures directory is empty.
4. Pin two C4-regression assertions onto the recorded ceo-bootstrap-empty
fixture: the assistant response must NOT contain "node -e" OR
"/api/companies". Any future regression that reintroduces the
HTTP-path-in-prompt pattern or the bash+node -e shell fallback will
now fail pnpm test:eval instead of living in the fixture silently.
typecheck green, test:eval up from 8 → 10 tests (2 new self-tests), all
passing.
Closes review asks from cross-review of C4:
- Jobs ask #2: team-building.en.md still referenced the bash+fetch
pattern for goal creation. Rewritten as a deferred-tool note pointing
to C5 follow-up, without re-teaching the HTTP shape C4 just removed.
- Karpathy ask #1 (partial): re-recorded manager-task-handoff against
main to produce post-C4 positive evidence. VP Engineering has
hire_agent available via canRequestHire, but when directed NOT to
hire, the recorded trace shows task decomposition only (no
hire_agent invocation). 455 in / 769 out tokens; assertions updated
to pin /api/task/create + agent_ic_ slugs + p99 substring + max 3
turns. The `non-ceo-dm-reply` re-record is deferred to the follow-up
ticket.
- Karpathy asks #2, #3, #4 + Jobs asks #1, #3, #4: documented in
docs/proposals/C4-followup.md with concrete completion criteria.
Scope covers canFileCeoStrategy predicate, non-ceo-dm-reply
re-record, RFC 8594 Deprecation/Sunset headers + counter, and a
negative substring assertion kind for the eval harness.
Eval still green at 8/8. C4 main batch (e2c839b..8e99e01) stands
unchanged; this commit is pure follow-up groundwork.
This is the verification artefact for the C4 refactor (e2c839b through
03864f3). The old fixture (recorded at 1d0cfa2, scenario asked the model
to POST hire_agent requests via bash+node+fetch) captured the model
emitting `node -e 'fetch(\`http://...\`)'` — exactly the antipattern
Jobs called out.
Post-C4 scenario was updated to reflect the new world: the CEO now has
explicit hire_agent and file_approval tools. Re-recording against
anthropic/claude-sonnet-4-5 produced 509 in / 731 out tokens and the
assistant response:
- Contains `hire_agent` (pseudo-XML tool invocations, one per hire)
- Contains `file_approval` with type=approve_ceo_strategy
- No `node -e` / bash+fetch / HTTP POST pattern anywhere
- 3-5 hires as requested, single approval block
Assertions updated to pin the post-C4 tool names as positive signals.
Absence of bash/fetch is demonstrated by diff against 1d0cfa2 (see that
commit's recorded JSONL for the before state).
Eval harness now green at 8/8 (3 self-tests + 5 recorded fixtures) with
this new baseline. Any future regression that reverts CEO bootstrap to
HTTP-path prompts will fail all three hire_agent / approve_ceo_strategy /
file_approval assertions.
These are the role×trigger tripwires Karpathy required before C4 can
ship: covering IC / manager / non-CEO paths, not just the two CEO
fixtures from 1d0cfa2. Together they act as the cross-role regression
net — any C4 change that breaks hire_agent tool registration or
prompt shape across role tiers will now surface as an assertion
failure, not a silent production drift.
Fixtures recorded against anthropic/claude-sonnet-4-5 via the user's
Claude Max OAuth token (read from macOS keychain, not committed):
- ic-task-execute (364 in / 1425 out): IC receives task t_1138,
expects task id + filename in the response and a mark_complete
intent within 3 turns.
- manager-task-handoff (457 in / 1022 out): VP Engineering decomposes
CEO p99 directive into IC-sized tasks; expects /api/task/create
calls, references to agent_ic_* slugs, and the "p99" term.
- non-ceo-dm-reply (357 in / 545 out): VP Product handles a peer DM
with blocking ETA ask; expects mark_read + /api/agent_message/send
+ message id preserved.
Total spend ~3.7k tokens; pnpm test:eval --run now green at 8/8
(3 harness self-tests + 5 recorded fixtures).
With these tripwires in place, C4 (hire_agent / file_approval tool
uplift + initial-user-turn rewrite) can proceed without the previous
"only CEO path has a baseline" blind spot.
Unblocks the march-of-nines third nine. Up until this commit `tests/eval/`
had a harness + self-tests but zero real traces; the fixture directory was
a documented stub. These two fixtures are genuine single-turn recordings
against anthropic/claude-sonnet-4-5 produced by `pnpm eval:record` — not
mock data, not placeholder — each one captures an actual API response
with its token ledger.
Coverage:
- ceo-bootstrap-empty: empty company, CEO bootstrap trigger. Asserts the
model produces hire_agent POSTs, an approve_ceo_strategy POST, and uses
bash+fetch (current prompt pattern before C4 lands).
- ceo-board-message-reply: operating CEO with a live team, inbound Board
message. Asserts the model hits mark_read, posts to
/api/agent_message/send, and keeps the original message_id in scope.
What these fixtures now enable:
- B2 Phase 2 switch-read: a fixture diff before/after the cutover
demonstrates dual-write parity is real, not assumed.
- C4 initial-user-turn refactor: the current HTTP-path-in-prompt pattern
is captured; after C4 lands, the diff shows whether tool-shape
refactoring regressed behaviour.
- C7 prompt-builder conditional assembly: same, but for system-prompt
slimming.
- M7 Phase 2 call-site switch: resolveCeoMode cutover can be regression-
verified against both traces.
Scenarios live at tests/eval/scenarios/<id>/{system.txt,user.txt} so any
maintainer can re-record the same trace with one command and diff against
the committed fixture. Each fixture's assertions block currently contains
four assertions; pnpm test:eval passes with 5 (3 harness self-tests + 2
real-fixture runs).
Replaces the stub recorder with a real-provider CLI. The command now:
1. Reads system prompt + user message from files on disk (empty files
are rejected — a fixture with no prompt captures no behaviour).
2. Resolves the provider spec from `--provider anthropic|openai|google`
and pulls the API key from the matching environment variable.
Missing env → hard error with a copy-pasteable export hint; no
fallback, no mock.
3. Calls `streamSimple` from @mariozechner/pi-ai with the user-picked
model. On provider error (stopReason=error|aborted) the CLI exits
non-zero and writes no JSONL. Users re-run if they think the
failure was transient.
4. Serialises one user turn + one assistant turn + an empty
assertions block into the JSONL shape documented in
tests/eval/README.md. Token ledger comes straight from the
assistant message's usage field (renamed to snake_case for
provider-neutral on-disk format).
Scope is deliberately single-turn: tool use / multi-turn agent loops
are B5.1's job. Importing the AgentCompany worker's system prompt
builder would require the Electron runtime; the CLI stays node-only
so it can run standalone.
`--help` works and prints the argument matrix. All failure modes exit
non-zero with a specific, actionable message.
Also document the new recording workflow in tests/eval/README.md —
four-step procedure, example scenario files, assertion templates, and
explicit list of what the recorder still does not do.
- harness.test.ts: add in-memory self-test layer (pass + fail + purity), guard
fixtures dir with existsSync so absent dir yields empty list without
crashing the runner. Empty-fixture pnpm test:eval now carries real signal.
- replay-client.ts: remove dead if (!turn) branch; cursor bounds-check above
already guarantees the element. Use non-null assertion with a comment.
- assertion-judge.ts: drop unnecessary (turn as AssistantTurn | { content:
unknown[] }) cast. turn.role === assertion.role already narrows to
UserTurn | AssistantTurn and both expose .content.
- _BLOCKED-BY-CONSTRAINT.md: correct electron-builder note. files: is a
whitelist, so tests/eval is automatically excluded and no new rule is
required.
Land the B5 harness skeleton: fixture loader (strict JSONL contract with
meta header + ordered turn records + assertions tail), replay-client that
swaps in for pi-ai's streamSimple by replaying recorded assistant turns,
structural assertion judge (tool_called / no_tool_called /
message_contains_substring / tool_call_arg_matches / trace_max_turns), an
eval-runner that sums the token ledger and produces an EvalResult, a
vitest driver (tests/eval/harness.test.ts) that discovers *.jsonl
fixtures, and a record-fixture CLI that writes fixture meta headers.
Hard contracts:
- No fallback: replay-client and fixture-loader throw on exhaustion /
malformed input rather than synthesizing defaults.
- No mock: types reuse @mariozechner/pi-ai AgentMessage / ToolCall
verbatim; shadow types are forbidden.
- No fake fixture data: tests/eval/fixtures/ ships only README +
.gitkeep. Real traces are recorded by the user against live providers.
package.json adds test:eval (vitest --passWithNoTests so an empty
fixture dir is green) and eval:record (CLI entrypoint). electron-builder
already excludes tests/ via its files: glob — fixtures never enter the
shipped .asar.
ZacharyZhang-NY