Multi-harness support: Codex + agents-md targets, target-matrix contract suite, library workflow, flag-drift canary - #9

One command, two tiers, no hostages. Without a Pro key, `bobby app`
(and its alias `bobby dashboard`) serves the classic dashboard exactly
as published — free forever. With Bobby Pro, the same command serves
the App, with classic kept at /classic/.

The App UI moves out of this MIT package into @bobbycode/pro-dashboard
(private), delivered via `bobby pro install` — a license check on
MIT-licensed files would be theatre, so the paid code simply never
ships here. commands/app.js resolves the UI: BOBBY_APP_DIR (dev) →
Pro key + installed package → classic fallback, and prints which tier
it chose and why.

The loop API (brief/go/ticket writes/workflows/config) stays MIT — it
serves the classic UI and the relay too.

Verified all three tiers live: no key → classic + upsell note; dev
override → App; real signed key + `bobby pro install <tarball>` from
the actual buyer zip → App at /, classic at /classic/. 855 tests green.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Approach A (foundation-first) selected 36/29/19: matrix invariant suite
first, then Codex target + executor, then generic agents-md target, docs
last. Scoring rationale, spec, and the no-unverified-harness-claims rule
in the epic's plan.md.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The Pro package now carries both the Bobby App (app/) and the classic
dashboard add-on assets (ui/), so the App lookup moves to app/.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…Zed, Gemini hold

Six more children under the multi-harness epic, all demand-driven and
gated on the matrix suite + agents-md generic tier landing first.
Windsurf/Zed are spikes where "covered by generic" is a valid close;
Gemini/Antigravity is an explicit dated hold, not a build.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The proxy test slept a fixed 50ms for a real HTTP round trip, which
lost the race under full-suite parallel load and failed intermittently.
It now resolves on the frame itself.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

bobbycode has no live app, so the built-in default workflow's final test
stage (live-app verification by contract) would strand every ticket at
testing. Override default/secure to end at review locally, define real
feature areas, and file TKT-013 (critical) for the upstream fix: built-in
library workflow, stack-aware default, and a loud block instead of a
silent stall. Also notes the workflow-list display bug for overridden
workflows.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Per-PR verification proves argv correct at merge time only; harness CLIs
ship weekly. A scheduled workflow probes each executor's real buildArgs
output against the real binary: auth failure = pass, unknown-option =
drift, plus a control probe so a CLI that ignores unknown flags can't
make the canary vacuous. TKT-004/TKT-009 now require registering their
flavors in it.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

One contract, run against every target in the registry via
describe.each(TARGETS). Adding a target now inherits ~all coverage from
registration alone instead of needing a hand-written suite.

Invariants: adapter interface shape, all declared paths written, extras
created, no writes into another target's directories, no scaffolded file
referencing another target's rules file, rules names its own harness,
skills/agents reference own paths, hooks scaffolded only where supported,
idempotent rescaffold, pre-existing rules backed up and merged, command
frontmatter matches the adapter's own transform, prompts reference only
this target's agents path.

Verified the suite catches the bug it exists for: reintroducing the
literal `CLAUDE.md` reference in templates/agents/bobby-build.md.ejs fails
the cline and cursor legs while claude-code (whose rules file that is)
correctly passes. That bug shipped in Cline for its entire life because
only cursor ever got a hand-written sweep test for it.

Removes 262 lines of per-target assertions the matrix now covers; keeps
genuine quirks (cursor transformCommand edge cases, ignore-file contents).

Refs TKT-002

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Target: rules -> AGENTS.md, skills -> .codex/skills, agents ->
.codex/agents, prompts -> .codex/prompts.

Executor: `codex exec --json`, derived from target: codex.

Everything verified against the real @openai/codex 0.146.0 binary
(strings + `codex exec --help`), never documentation:

- AGENTS.md is Codex's native project-instruction mechanism — the binary
  carries a full "AGENTS.md spec" in its base instructions and ships
  /init to create one.
- The prompt is POSITIONAL. `-p` is `--profile` in Codex, so reusing the
  claude/cursor `-p <prompt>` shape would have silently passed the whole
  prompt as a config profile name with no error. The prompt is emitted
  last so it can never be read as a flag value.
- JSONL comes from `--json`; there is no --output-format and no --verbose.
- Permission modes map to verified flags:
  bypassPermissions -> --dangerously-bypass-approvals-and-sandbox
  acceptEdits       -> --sandbox workspace-write --ask-for-approval never
  plan              -> --sandbox read-only

Correction to the ticket's premise: Codex skills are USER-level
($CODEX_HOME/skills, default ~/.codex/skills). Project-level auto-loading
of .codex/skills/ is NOT confirmed, so the loop does not depend on it —
AGENTS.md names the directory and every prompt references its agent and
skill by path. supportsSubagents() is false for the same reason: the
binary mentions subagents but the definition format is unverified.

Live-verified: Bobby's exact argv run against the real binary starts a
turn and streams JSONL (thread.started / item.completed / turn.started),
with no unknown-option error.

Codex passes all 19 target-matrix invariants with zero test edits beyond
registration — the matrix suite (TKT-002, cherry-picked here to prove it)
went 47 -> 66 tests on target count alone. 914 total green.

Refs TKT-003, TKT-004

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Rules -> AGENTS.md, skills -> .agents/skills, agents -> .agents/agents,
commands -> .agents/commands. Lets Bobby work in any tool that reads the
AGENTS.md convention without shipping a dedicated adapter for each.

Verified against shipped code, not documentation:
- Cursor 3.13's skill-root array literally contains ".agents/skills/",
  alongside .cursor/, .claude/ and .codex/ skill roots.
- @openai/codex 0.146.0 carries a full "AGENTS.md spec" in its base
  instructions and uses the adjacent ".agents/plugins" namespace.

Scope is deliberately narrow and asserted by tests: rules + skills only.
No subagent claim (no cross-tool convention exists; prompts reference
agents by path). No dashboard executor derivation — unlike cursor/codex
this target does not imply a specific CLI, so the dashboard stays on
claude unless dashboard.executor says otherwise.

Passes all 19 target-matrix invariants with zero test edits beyond
registration. 906 tests green, lint clean.

Refs TKT-005

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Documents all five targets — claude-code, cursor, codex, cline, agents-md
— with rules/skills/commands/agents paths, subagent support, and which
dashboard executor each derives.

The second table is the point: it records HOW each row was checked, not
just what it claims. Bobby has shipped wrong harness claims before (a
model name copied from a CLI's own stale --help, a "no subagent registry"
line a newer build contradicted), and both came from trusting docs. So
rows are marked Live / Live argv / Paths verified / Unverified, and cline
is explicitly labelled best-effort rather than quietly implying parity.

Also states the agents-md tier's limits plainly (rules + skills; no
subagent dispatch, no executor derivation) and adds a contributing
section pointing at cursor.js as the template and the matrix suite as the
acceptance bar, with the no-unverified-claims rule.

Verified: a script cross-checks every documented path, subagent flag, and
executor derivation against the actual adapters — all five rows match
shipped behaviour. 936 tests green across 48 suites.

Refs TKT-006

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Per-PR verification proves an executor's argv is correct at merge time and
nothing after; these CLIs ship weekly. A scheduled workflow (Mondays, plus
manual dispatch — never on push/PR) probes the real binaries with no
account required.

Mechanism: agent CLIs parse argv before checking auth, so Bobby's exact
command line run unauthenticated must fail on auth, never "unknown
option". A second control probe adds a bogus flag that MUST be rejected —
without it, a CLI that silently ignores unknown flags would produce a
vacuous pass.

argv comes from the real buildArgs functions. Verified by the AC: zero
flag strings appear in the workflow file, so the canary cannot drift from
the code it guards.

Building this caught a bug in the canary itself: resolveExecutor treats an
unrecognized name as a custom binary path with claude-style flags — correct
for `dashboard.executor: /path/to/x`, catastrophic here, since probing an
unregistered flavor built claude argv and reported it as codex drift.
Now guarded against EXECUTOR_NAMES with a distinct unknown-flavor outcome.

Live-verified against both installed CLIs: codex stops at authentication
as expected; cursor-agent clears argument parsing. A missing binary
reports install-failed, distinct from drift.

Failures file one idempotent GitHub issue per flavor (commenting on the
existing one rather than piling up weekly duplicates).

Refs TKT-014

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

CLIs, libraries, and npm packages silently stalled at the `test` stage.
bobby-test verifies by exercising a running application and its skill
forbids running the spec suite as a substitute, so on a project with
nothing to serve every test case came back BLOCKED and the ticket simply
stopped advancing — no error, no guidance.

- Built-in `library` (plan/build/review) and `library-secure` workflows.
  Review already runs the suite independently, which is the right
  verification for a library.
- `default_workflow` config key, and `stackDefaultWorkflow()` which honors
  an explicit `default_workflow` in stack JSON and otherwise infers
  `library` when a stack has no health checks and no dev command. `bobby
  init` writes the choice into .bobbyrc.yml with a comment explaining why,
  so it is visible rather than implicit.
- Live-app agents (test, qe, ux, pm, watchdog) invoked where there is
  nothing to observe now block the ticket with an actionable reason and
  name the fix, instead of emitting BLOCKED cases one at a time.
- `bobby workflow list` shows effective stages for overridden workflows
  (it printed the shadowed built-in ones) and marks the project default.

Dogfood proof: bobbycode's own .bobbyrc.yml drops its manual override for
`default_workflow: library`, and the loop still ends at review.

Refs TKT-013

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>