Bring the grok CLI to parity with the Claude path on the four deliberate
differences:
- Reasoning: drop the per-role `--effort low` default — Claude sets no per-role
thinking budget, so grok now uses the model default for every role. The
fleet-wide ROBOCO_GROK_REASONING_EFFORT override stays as a cost lever. (This
also un-caps intake-draft quality, the one that actually mattered.)
- Subagents: the intake interviewer may now fan out to subagents (parity with the
Claude intake's `Task` allowance); every other role still has `Agent` removed.
- Web: `--disable-web-search` for every role — no agent gets direct web (Claude's
tool set has none either); the roles that get web reach it through the gated
roboco-search MCP, unaffected.
- Bash command filtering: full parity, split by deny semantics. Verified live that
a grok PreToolUse hook deny CANCELS the run, while native `--deny` denies
GRACEFULLY (the agent gets a permission error and recovers). So:
* git network/branch/history ops -> native `--deny` (operational reflex; the
agent must recover, not drop the task). Expanded to the full bash-guard set.
* credential-exfil / identity-forgery / internal-API / env-dump patterns ->
the SAME bash-guard the Claude path runs, wired as a grok PreToolUse hook
(ROBOCO_GUARD_SKIP_GIT=1 so it leaves git to `--deny`). A hard cancel is the
right response there — no legitimate agent reads ~/.netrc or forges an
X-Agent-ID. One tolerance line (accept grok's camelCase `toolInput`) makes
the one tested script guard both runtimes; +5 grok cases (50/50 green).
Also cleaned stale internal task-number / smoke labels out of bash-guard-hook.sh.
The blueprint was mounted at /app/system-prompt.md but never reached grok — a real
parity gap vs the Claude path (which passes --system-prompt-file). grok agents ran
only on the per-task prompt, missing their RoboCo role/org context.
Verified live on grok 0.2.56 that the obvious flags do NOT work headless:
`--system-prompt-override` and `--rules` are silently ignored under `grok -p`
(identical output with and without). What IS honoured is grok's instruction-file
discovery — and `$HOME/.grok/AGENTS.md` is loaded GLOBALLY regardless of --cwd
(a project AGENTS.md only loads from the cwd/project root, which would pollute the
agent's git workspace). Proven end to end: a blueprint written there makes grok
adopt the role ("I am the RoboCo intake interviewer ... -- intake-1").
write_agents_md() copies /app/system-prompt.md -> ~/.grok/AGENTS.md; the one-shot
render (grok_cli_config.main) and both interactive mains call it. No git pollution
(it lives in ~/.grok, not the workspace), and it covers repo-cwd and /app-cwd
roles alike. Reverted the non-working --system-prompt-override wiring.
Move the last Grok runtime off opencode onto xAI's official `grok` CLI, for full
parity with the Claude path. The intake/secretary chat now runs per-turn headless
`grok -p` invocations that resume one session id (proven live: context carries
across runs), with streaming-json deltas mapped to the existing panel StreamChunk
kinds — the IntakeDriver loop, message source, relay, and idle reaper are reused
unchanged; only the SessionFactory differs (GrokCliSession replaces the
opencode-serve session).
- GrokCliSession + a pure, unit-tested streaming-json -> StreamChunk assembler
(thought coalesced to one block, text streamed live, end captures the session
id for -r, fenced-draft fallback, clear errors incl. rate-limit).
- intake propose_draft and secretary read_company_state/read_task/submit_directive
are now FastMCP servers (roboco-intake / roboco-secretary) wired into
~/.grok/config.toml, launched via `uv run --directory /app` to resolve the
installed package. The secretary tools reuse the shared backend helpers.
- Orchestrator: interactive spawn mounts the subscription auth + per-agent usage
dir (no metered xAI key, no permission env — grok flags carry per-role perms);
usage/cost now read a captured usage.json (drop the opencode.db reader, the
_opencode_db_path/_grok_usage_from_opencode methods, and the cost-cap's
opencode read). hosts["opencode"] -> hosts["grok_usage"]; OPENCODE_DATA_DIR ->
GROK_USAGE_DATA_DIR.
- Fix one-shot usage capture: `-s` does not pin the session id (grok generates
its own), so the entrypoint now reads the real id back from the JSON run log
and the reader uses it; usage is captured per-turn on the interactive path.
- Delete the opencode layer: opencode_config/opencode_usage/opencode_session, the
docker/grok/*.js plugins, the old one-shot entrypoint, and their tests.
- Compose (all three files), .env.example, and stale comments updated to the
grok-CLI runtime; add the SuperGrok auth mount + grok-usage dir.
Gate green: ruff, mypy (296 files), xenon, tests. NAS build/verify pending.
The provider mounts the per-agent data dir and points the entrypoint's usage
file at it; the orchestrator's grok finalize reads that usage.json first (the
grok-CLI total, priced at the output rate) and falls back to opencode.db for the
still-opencode interactive intake/secretary path. Re-add _reasoning_effort_for to
grok.py as a clearly-temporary shim for that interactive path (it needs opencode's
"minimal" variant, distinct from the CLI's --effort) until it is converted too.
The provider pins a fixed session id (ROBOCO_AGENT_SESSION_ID, reused from the
agent session id as on the Claude path); the entrypoint passes it to
'grok -p -s <id>' so the run's session store is locatable, then runs the usage
reader post-run (best-effort) to write the captured tokens + cost. The
orchestrator-side finalize that reads that file follows.
Grok runs on the SuperGrok subscription, but — exactly like Claude on Max — we
still record per-agent tokens and a notional cost for the dashboard. The grok
CLI writes a cumulative totalTokens per turn into
~/.grok/sessions/<cwd>/<session-id>/updates.jsonl (the grok analogue of the
Claude transcript / old opencode.db); the max is the session total. This reader
locates that file (url-encoded cwd), extracts the total, and prices it at the
output rate (no input/output split from the CLI; conservative + matches the
reasoning-at-output convention). Validated against a real grok-build session
(18253 tokens -> $0.0365). Entrypoint + finalize wiring follows.
Replace the opencode GrokProvider with GrokCliProvider: reuses the orchestrator's
shared mount/auth/git assembly (gateway + identity) exactly like the Claude path,
mounts the host ~/.grok/auth.json read-only (SuperGrok subscription) instead of
injecting an xAI key, and sets the slim env the grok-cli entrypoint + renderer
read (ROBOCO_AGENT_ID for per-role flags, model, mcp-config, prompt). Provider
routing fields are blanked before the shared step so the grok endpoint is never
mislabelled ANTHROPIC_*. Per-role permission logic now lives in grok_cli_config,
so the provider is slim. Registry/orchestrator/exports updated; provider tests
rewritten for the CLI behavior (no XAI key, auth mount present/absent).
First piece of the Grok CLI provider that replaces the opencode runtime: a
pure, unit-tested module the agent entrypoint runs to translate the mounted
mcp-config.json into ~/.grok/config.toml ([mcp_servers]) and compute the
per-role 'grok -p' flags — subagent/shell/edit tool removal, raw-git-mutation
and rm-rf denies, reasoning effort — mirroring ClaudeCodeProvider's per-role
permissions with native grok flags instead of an opencode permission block +
JS guard plugins. Uses tomli_w. The rendered config + env injection are
validated live against grok-build (the model called the server through it).
Three bugs surfaced by the first live Grok lifecycle run:
- Dev/QA/doc agents crash-looped at startup with ModuleNotFoundError on
roboco.llm.providers. The entrypoint ran the opencode-config render from the
agent's workspace-clone cwd, whose own roboco/ dir shadows /app on the
sys.path front; a branch without the grok code lacks the providers package.
Render from /app so the installed package always resolves (the render has no
cwd dependency — writes global, reads ROBOCO_MCP_CONFIG).
- A budget/loop halt blocked EVERY tool, including i_am_idle, unclaim, and
i_am_blocked, so a halted agent could neither continue nor stop and flailed —
one billed model turn per blocked retry. The before-gate now always lets the
release verbs through so a halted agent can exit cleanly.
- The inbound reviewer ingested the org's OWN PRs (authored by the repo-owner
account), which can't take a REQUEST_CHANGES review (GitHub 422) and get
re-reviewed every poll. The normalizer flags author_is_owner and ingestion
skips them — the reviewer reviews only PRs the org did not author.
External/contributor PRs are unaffected.
Tests: owner-authored PR flagged + skipped; normalize shape covers the new
field. Gate green on the changed modules (ruff/mypy/xenon + 48 tests).
A pr-reviewer review of an org-authored PR never reached GitHub. The agent
side ran correctly (claim → read-only diff → review → post_pr_review →
completed + CEO notify), but the GitHub publish 422'd with "Can not request
changes on your own pull request": the PR was authored by the same account
that owns the project PAT. post_pr_review posts best-effort after the DB
transition, so the failure was logged and swallowed — the task completed and
the CEO was notified "reviewed" while the PR showed no review.
GitHub forbids APPROVE / REQUEST_CHANGES on your own PR but DOES allow a
plain COMMENT review. The org's internal PRs (and any PR the PAT owner
opened) hit this. Retry once as a COMMENT review on the self-review 422 so
the review actually lands; the verdict is already stated in the body. The
external/fork-PR path (different author) is unchanged — REQUEST_CHANGES
succeeds there and the fallback never fires.
Tests: self-review 422 downgrades to COMMENT and returns the COMMENT result;
a failing COMMENT retry still surfaces GitError with no infinite loop; the
existing non-self 422 still raises.
An interactive intake/secretary chat the human abandoned (closed the tab without
confirming or stopping) leaked its container until the orchestrator restarted —
the wedged-grok reaper is task-driven and these run task_id=None, and an SSE
disconnect intentionally does NOT reap (so a page reload can reconnect).
Reap by IDLE TIME, not connection state: PrompterLiveRegistry tracks
last_activity (bumped on every push/deliver = a turn), and the 60s sweeper
retires sessions idle past ROBOCO_INTERACTIVE_IDLE_REAP_SECONDS (default 1800;
0 disables) via reap_intake_session / reap_secretary_session. An active or
page-reloaded chat that keeps exchanging turns stays fresh and is never reaped;
board-review-parked sessions (task_id set) are exempt. Provider-agnostic — fixes
the leak for both Claude and Grok interactive.
Tests: idle-only reap (active/parked/closed excluded), activity bump keeps a
session alive, threshold 0 disables. Gate green (ruff/mypy/xenon + prompter_live).
Live verification (opencode 1.17.8 + grok-build-0.1, funded key) showed the Grok
runtime was loading INERT, three ways:
1. The provider override `provider.xai.npm=@ai-sdk/openai` failed model
resolution (ProviderModelNotFoundError) — opencode can't resolve that package
from its module path. Worse, ANY custom `provider.xai` block (even just
options) breaks plugin-tool registration. opencode's BUILT-IN xai provider
drives grok-build-0.1 with working tool-calls, so emit NO provider block; the
key + base reach it via XAI_API_KEY / XAI_BASE_URL env (provider.options.apiKey
alone does NOT authenticate).
2. Plugins referenced by absolute path in the config `plugin:` array never
registered their hooks/tools. opencode 1.17.8 only registers from the plugin
AUTO-DISCOVERY dir (~/.config/opencode/plugin/). Bake all plugins there.
3. Plugins must use a NAMED export, not `export default`.
Changes:
- opencode_config: no `provider` block, no `plugin` array; drop the dead
XaiTarget + timeout machinery; build_opencode_config now takes a model string.
- GrokProvider / orchestrator interactive env: inject XAI_API_KEY + XAI_BASE_URL
(drop the now-unused OPENAI_*).
- secret-scrub / budget-feed / secretary-tools / intake-tools: named exports;
baked into /home/agent/.config/opencode/plugin/ (drop the EXTRA_PLUGINS env).
- agent-grok* Dockerfiles: plugin dir + agent ownership; drop the unneeded
@ai-sdk/openai global install.
Verified live end-to-end: grok-build-0.1 calls read_company_state AND
submit_directive through secretary-tools.js and the backend receives both with
the agent token; a tool.execute.before guard fires; built-in tool-calls work.
Targeted gate green (ruff/mypy/xenon + opencode_config/providers/interactive
tests; node --check the plugins).
_enforce_grok_cost_budget killed + evicted a container directly. For the
interactive roles (intake/secretary) that left the panel SSE relay open with no
close sentinel, so the chat froze with no explanation. Add
PrompterLiveRegistry.close_by_agent (push a final error event, then close every
session bound to that agent) and call it from the cost-cap watchdog when the
killed agent is the intake or secretary, so the panel reports the chat ended on
the cost cap instead of hanging.
If `opencode serve` died after the session opened, every subsequent turn failed
with an opaque httpx connection error while the container lingered. send() now
detects the exited subprocess (returncode set) and yields a clear error chunk +
turn_end so the panel shows a real "session ended — start a new chat" message;
the idle watchdog / a human reap then tears the container down.
A one-shot grok run that hit an xAI 429 exited without a terminal verb; the
dispatcher then re-spawned the same task every tick (429 -> exit -> respawn), a
container/token/cost loop with no living agent to call i_am_blocked.
- entrypoint: detect a rate-limit signature in the run output and exit 75
(EX_TEMPFAIL); a rate-limited task is NOT substituted — it must be retried.
- _handle_stopped_container: on a grok exit 75, park the provider via the
rate-limit tracker (retry_after window) instead of crash-retrying, and don't
count it as a crash. The existing probe-resume loop clears the park after the
window (unknown-provider time-expiry fallback) and the task is retried.
- spawn_agent: a grok-only, fail-open guard skips the launch while the provider
is parked, so the dispatcher no-ops instead of looping. The Claude path is
untouched.
Targeted gate green (ruff/mypy/xenon + new rate-limit tests; bash -n on the
entrypoint).
Grok wrote ONE global permission block, so a Grok pr_reviewer (or qa / PM /
auditor) ran with edit=allow + bash=allow on untrusted PR content. Now the
permissions are derived per role, mirroring orchestrator._get_role_permissions
on the Claude path:
- edit — allow only roles that write code (role_config.allows_write:
developer / documenter); everyone else edit=deny.
- bash — allow only roles that legitimately run a shell (developer /
documenter / cell_pm / main_pm); the read-only reviewers (qa / pr_reviewer /
auditor) and the board get bash=deny. secret-scrub still guards the rest.
- external_directory — only the pr_reviewer reads scratch outside its cwd (the
/tmp diff); delivery roles get deny.
One-shot roles resolve these in GrokProvider._append_grok_env; the interactive
intake/secretary set edit=deny + bash=deny in the orchestrator (intake keeps
external-dir reads for sibling product repos, the secretary does not). The
Claude path is untouched — the permission env is a GROK-only contract.
Targeted gate green (ruff/mypy/xenon + provider + interactive-spawn tests).
The prompter prompt tells the model to call propose_draft when the spec is
ready, but on Grok that tool didn't exist — so no draft chunk, no panel draft
card, and the human couldn't launch a task from a Grok intake chat.
- intake-tools.js: opencode plugin registering propose_draft via Hooks.tool;
the execute() only ACKs — the driver (OpencodeServeSession.normalize ->
_is_propose_draft -> _draft_from_tool_input) intercepts the tool CALL and
emits the `draft` chunk the panel renders.
- agent-grok-prompter.Dockerfile: bake the plugin, scoped to this image via
ROBOCO_OPENCODE_EXTRA_PLUGINS (delivery roles never draft).
- test: a propose_draft tool part normalizes to a draft chunk (not a tool_use).
The live tool-call -> draft-card path is flagged UNVERIFIED-LIVE for the NAS.
The keystone of the Grok parity work (CEO's "take Claude as baseline, create
what's missing" call): the one-shot Grok container now starts the same SDK
server the Claude path runs, so the per-verb circuit breaker (the flow/do MCP
servers already POST /verb/attempted to it), the per-session budget/loop
counters, the terminal-verb tracking, and the SessionEnd post-mortem all work
on Grok instead of being silently absent.
- entrypoint: launch roboco.agent_sdk.server (bare venv python, not `uv run`
which would re-sync the drifted clone lock and stall), wait for /health,
reset counters; run opencode WITHOUT exec so the script regains control to
run the post-mortem and the silent-exit substitute after the run returns.
- budget-feed.js: opencode plugin that gates on /budget/status in
tool.execute.before (halt/loop deny — the only place to stop a runaway
one-shot run; opencode has no PostToolUse-deny) and records the executed
tool + args-hash in tool.execute.after. Fail-open; bare-verb normalization
for MCP-namespaced terminal verbs.
- silent-exit substitute: on a graceful exit with no terminal verb the
entrypoint posts /terminal/force_substitute so the task isn't left stuck
claimed/in_progress (Stop-hook parity at the boundary).
- opencode_config: wire budget-feed into the plugin array; add
ROBOCO_OPENCODE_EXTRA_PLUGINS so per-image role tool plugins load scoped to
one role; read the per-role ROBOCO_GROK_EDIT_PERMISSION.
Targeted gate green (ruff/mypy/xenon + opencode_config tests; node --check on
the plugins; bash -n on the entrypoint).
Live NAS run showed the Grok pr-reviewer claim the review and fetch the diff,
then write it to /tmp and FAIL to read it back: opencode auto-denied
"external_directory (/tmp/*)" — its file tools refuse paths outside the project
cwd, and in headless serve/run mode an "ask" permission auto-rejects (no human).
Add permission.external_directory (default "allow", env
ROBOCO_GROK_EXTERNAL_DIR_PERMISSION) to the generated opencode.json. The
container is the sandbox and secret-scrub still blocks credential-file reads, so
allowing in-container external-dir reads is safe and unblocks legitimate scratch
use (e.g. the pr-reviewer grepping a large diff in /tmp).
Verified live against grok-build-0.1: with external_directory:"allow" the Read
tool reads a file outside cwd and returns its contents (no auto-reject); the
plain-string form is accepted by opencode 1.17.8.
Needs a rebuild of roboco-agent-grok + a pr-reviewer re-run on the NAS to confirm.
The injection guard is RoboCo's own hook (user-prompt-hook.sh), not a runtime
built-in, so it can be recreated at our input boundary regardless of runtime —
opencode's lack of a blocking pre-prompt hook is irrelevant.
- prompt_guard.detect_injection: the deny patterns ported to reusable Python.
- IntakeDriver._run_turn scans every interactive turn before sending it to the
model and denies a match as an error chunk. Covers BOTH Grok (opencode) and
the Claude SDK intake (which runs with setting_sources=[] and so never loaded
the bash hook — it was unguarded too).
- The one-shot grok entrypoint scans ROBOCO_INITIAL_PROMPT and refuses a
poisoned task prompt (parity with the Claude UserPromptSubmit deny).
- Broadened the pattern (Python + the bash hook, kept in sync) to catch the
multi-qualifier canonical phrasing "ignore all previous instructions", which
the single-qualifier original missed — without false-positiving on
"ignore the linting rules" (an intermediate non-qualifier word breaks it).
So Grok now has the command/secret-exfil guard (secret-scrub), the cost cap,
AND the injection guard. Verified: 94 agent_sdk tests pass; bash + Python agree
on detect/miss cases.
Found by actually running opencode serve locally (the path was doc-verified but
never executed). Three fixes:
1. EACCES on the opencode store mount (the live intake crash): on Linux docker
auto-creates a missing bind source as root:root, so the non-root agent user
could not mkdir/write in /home/agent/.local/share/opencode and opencode died
at boot. _ensure_opencode_data_dir pre-creates the per-agent dir 0777 before
the mount (one-shot via the _GrokHost seam, interactive in both spawns).
2. Silent blank reply on a model error: opencode reports a turn failure in
info.error with parts=[], NOT as a part — verified live (a bad xAI key
returns info.error APIError). send() / normalize_opencode_message now surface
it as an "error" StreamChunk so a failed turn is never blank (the original
intake bug class). Confirmed live: the error now renders.
3. Reasoning variant on the serve path: the live OpenAPI shows the message body
accepts a "variant" field (it is NOT CLI-only, as the docs implied), so the
pin is unblocked. send() passes ROBOCO_GROK_VARIANT as the per-turn variant;
the orchestrator sets it per-role (_reasoning_effort_for) for interactive
Grok, the same lever as the one-shot --variant.
opencode serve startup, POST /session, session-id extraction, the part-type
mapping (text/reasoning/tool), and the error path are all validated against a
live opencode 1.17.8. A real successful grok reply still needs a funded key.
Claude Code's per-agent token-budget hook fires against the SDK :9000 server;
opencode exposes NO usage/budget hook to a plugin (confirmed against its plugin
docs), so the budget kill-switch can't be a plugin/sidecar — the orchestrator
enforces it instead.
_enforce_grok_cost_budget runs each dispatch tick: for every ACTIVE GROK
container it reads cumulative cost from the opencode store (the Phase-2 reader)
and kills + evicts it past ROBOCO_GROK_MAX_COST_USD (0 = off), after which the
reaper releases the freed task. This also catches a runaway loop that keeps
firing verbs (so it evades the idle watchdog) but still burns cost.
Covers the budget/runaway-burn slice of guardrail parity. The remaining Claude
hooks (prompt-injection PRE-gate, stop-guard terminal-verb) have no blocking
opencode equivalent — opencode's message/stop hooks are observe-only — and the
interactive reasoning-variant has no opencode.json/serve knob (CLI-flag only);
both are pinned for a live probe rather than shipped as a guess.
The CI mypy scope (roboco/ tests/) flagged test-only typing issues my per-file
runs missed: direct method assignment (orch._remove_container = AsyncMock())
trips [method-assign], and a module-level dict[str,str] is invariant against
the dict[str, str|None] the run-spec expects.
- Use monkeypatch.setattr for _remove_container in the watchdog tests.
- Annotate the shared _HOSTS as dict[str, str | None].
Production code unchanged; mypy roboco/ tests/ is green.
Wire the GROK interactive path the in-place way (matching how the interactive
roles already choose ANTHROPIC_* per route), so a GROK route launches the
Grok-native opencode-serve image instead of the Claude SDK-driver image:
- _spawn_intake_container / _spawn_secretary_container pick the
grok-prompter / grok-secretary image (ensuring the base→grok→interactive
build chain) when the route is GROK, and stamp provider_type on the spec +
AgentConfig so finalize routes usage to the opencode store.
- _build_intake_run_cmd / _build_secretary_run_cmd inject OPENAI_* + the
opencode store mount + system-prompt env for GROK via a shared
_append_interactive_provider_env, keeping ANTHROPIC_* for every other
provider. The intake's minimal mounts (no gateway MCP) are preserved, so
Grok intake matches the Claude intake's tool surface (the spec).
- Add a per-agent opencode store mount to the interactive host paths so
interactive Grok usage/cost is captured like the one-shot path.
Removes the interim Phase-0 routing guard (the real path supersedes it) and
retires the unused AgentProvider.spawn_interactive/InteractiveSpawnSpec seam —
the interactive roles have a bespoke assembly that the one-shot provider
surface doesn't fit, so the fork lives in their own builders.
UNVERIFIED-LIVE: end-to-end intake/secretary chat on Grok needs the stack up +
opencode serve confirmed against grok-build-0.1.
Builds the Grok analogue of the Claude intake/secretary live-session runtime,
satisfying the same IntakeSession seam so the existing IntakeDriver loop,
message source, relay, and StreamChunk panel contract are reused unchanged:
- OpencodeServeSession: a held-open `opencode serve` session (context persists
across turns) where each human turn is one synchronous POST /session/:id/
message; normalize_opencode_message maps the reply parts to text/thinking/
tool_use/draft/turn_end chunks (draft via a propose_draft tool part or the
fenced roboco-draft fallback). Doc-verified against opencode's server API.
- grok_intake_main / grok_secretary_main: container entrypoints mirroring the
Claude mains but yielding an OpencodeServeSession; they render opencode.json
(xAI provider + MCP + system prompt) first, then run the receiver + driver.
- roboco-agent-grok-prompter / -secretary images (FROM roboco-agent-grok) +
their builder services in all three compose files.
UNVERIFIED-LIVE: the opencode serve flow + exact Part schema + draft path need
a live run against grok-build-0.1 (the part mapping is defensive). The
orchestrator wiring that routes a GROK intake/secretary route to these images
is the next step (a design decision is open — see the handoff notes).
The AgentProvider ABC modelled only the one-shot lifecycle (spawn/stop/
health_check/remove), so the interactive intake/secretary roles could never
route through a provider. Add an opt-in interactive surface:
- supports_interactive class flag (default False).
- InteractiveSpawnSpec: the resolved AgentConfig + session id + role-specific
image + optional HMAC token — everything a provider needs without importing
orchestrator internals.
- spawn_interactive(spec): a non-abstract default that declines via
ProviderError, so every existing one-shot provider is unchanged.
Pure scaffolding — no provider opts in yet (GrokProvider flips the flag when
its interactive driver lands). Zero behavioural change.
A GROK agent runs opencode, not Claude Code: it has no SDK /usage/status
server and writes no Claude transcript, so _resolve_final_token_usage found
nothing and every Grok agent finalized at 0 tokens / $0 — the opencode_usage
reader existed but had no caller.
- Mount a per-agent opencode data dir ($DATA/opencode/<agent_id> →
/home/agent/.local/share/opencode) so opencode.db is captured, and mount the
same host dir into the orchestrator (/data/opencode) in all three compose
files so the finalizer can read it back — the opencode analogue of the
mounted Claude transcript.
- _resolve_final_token_usage branches on provider_type: GROK reads opencode.db
via opencode_usage (reasoning folded into output, billed at the output rate)
and skips the SDK/transcript path. A 0-token read logs a WARNING so a silent
mount failure isn't mistaken for a real zero-cost run.
- ROBOCO_OPENCODE_DATA_DIR overrides the in-orchestrator path for local runs.
intake (prompter) and secretary run a held-open chat session driven by the
Claude Agent SDK. GROK has no interactive runtime yet, and a GROK route for
those slugs would be spawned with the route creds injected as ANTHROPIC_*
against api.x.ai/v1 — the wrong protocol — producing a silent, empty reply
(the blank intake we observed).
Downgrade a GROK route for intake-1/secretary-1 to the Anthropic default with
a logged warning. The one-shot delivery roles route to GROK unchanged. This
guard is replaced by the real interactive fork once the opencode interactive
driver lands.
The heartbeat reaper deliberately skips a task whose assignee holds a live
ACTIVE container, so a Claude agent deep in a long edit/test cycle isn't
churned out from under live work. A wedged opencode container breaks that
assumption: it stays ACTIVE while firing no gateway verb, so its heartbeat
never advances and the live-instance skip would shield its task forever — the
exact way the Grok pr_reviewer parked in_progress.
Add a longer grok-idle kill threshold (ROBOCO_GROK_IDLE_KILL_SECONDS, default
900s, well past the stream chunk timeout). A GROK instance idle past it is
force-removed (its logs dumped to disk first) and evicted from the instance
registry, so the same reaper pass then releases the task. Only GROK runtimes
are eligible — a quiet Claude agent keeps the heartbeat-skip protection.
The Grok pr_reviewer wedged in_progress forever: opencode's default agent ran
with the subagent `task` tool enabled, spawned an Explore subagent on
grok-build-0.1 whose model call opened an SSE stream that went idle, and the
run hung with no timeout.
- Hard-disable opencode's subagent `task` tool in the generated opencode.json.
No RoboCo role uses opencode-internal subagents — work flows through the
gateway verbs — so removing the tool kills the hang trigger outright.
- Set provider.xai.options.timeout + chunkTimeout (operator-tunable via
ROBOCO_GROK_REQUEST_TIMEOUT_MS / ROBOCO_GROK_CHUNK_TIMEOUT_MS) as the
defence-in-depth backstop; chunkTimeout aborts an idle stream.
- Bundle the permission + timeout + subagent knobs into an OpencodeGuards
dataclass (keeps the builder under the arg-count gate).
- Drop the dead ROBOCO_AGENT_TOOLS spawn env (it had no consumer); opencode
tool restriction lives in the rendered config now.
grok-build-0.1 reasons heavily by default and reasoning bills at the output
rate (a live "say ok" call emitted ~300 reasoning tokens, ~85% of its cost).
Confirmed live that opencode's `--variant minimal` cuts reasoning ~54%
(298 -> 136 tokens, same prompt).
GrokProvider now picks reasoning effort by role: code-quality roles (developer,
qa, pr_reviewer) keep full reasoning; coordination / docs / board roles
(cell_pm, main_pm, documenter, product_owner, head_marketing, auditor, prompter,
secretary) run "minimal". It's passed to opencode via the entrypoint's
`--variant`. Operators can force one effort for ALL grok agents with the
ROBOCO_GROK_REASONING_EFFORT env (minimal | high | max, or default/full).
Tests cover the role map, the env override, and the spawn env wiring.
The Routing-mode toggle had Anthropic / Ollama / Self-Hosted / Mix but no way
to route the whole org to Grok. Add it end to end:
- backend: apply_mode("grok") + _apply_grok (GLOBAL default -> grok-build-0.1) +
derive_mode "grok" detection; ApplyModeRequest/ModeResponse accept "grok".
- panel: a "Grok" routing-mode card (between Anthropic and Ollama, gated on the
xAI key) + flipToGrok; a Grok group in the per-agent mix dropdown +
catalogGrokOnly + a grok ProviderBadge variant; the mix-save key check and
the AI-routing description now cover Grok.
- tests: integration derive_mode/apply_mode "grok" cases (+ grok provider row
in the fixture).
Gated: ruff + mypy clean; panel typecheck + lint clean.
A live opencode run against api.x.ai/v1 surfaced three real bugs:
1. Provider package — grok-build-0.1 is driven via the OpenAI Responses API
(opencode calls model.responses()). @ai-sdk/openai-compatible is
chat/completions only and errors "responses is not a function". Switch the
generated opencode.json provider + the grok image to @ai-sdk/openai.
2. Headless hang — `opencode run` blocks after init without a TTY; close stdin
(`< /dev/null`) in the entrypoint so it proceeds to the model call.
3. Reasoning-token cost — grok-build-0.1 is a reasoning model; reasoning tokens
bill as output but opencode stores them in a separate column. cost_for_session
folds tokens_reasoning into output (else ~22x undercount).
Verified end-to-end against a real session row (input=6120, output=1,
reasoning=226, cache_read=1856): our pricing reproduces opencode's stored USD
cost ($0.0069452) exactly. Tests anchored to that real row.
Confirmed by inspecting a local opencode run: opencode persists per-session
usage in SQLite at ~/.local/share/opencode/opencode.db — the `session` table
carries cost + tokens_input/output/reasoning/cache_read/cache_write. xAI's
response usage object (prompt_tokens, completion_tokens,
prompt_tokens_details.cached_tokens, completion_tokens_details.reasoning_tokens)
maps directly onto those columns.
Add opencode_usage.read_session_usage / cost_for_session: read the opencode DB
and price the tokens via roboco.billing.pricing (our cost stays authoritative;
opencode's own `cost` column is kept for reference). Tested against a fixture DB
mirroring the real schema (single session, summed sessions, missing/empty DB).
Remaining wiring (for the live spawn): mount the opencode data dir on grok
spawn + call cost_for_session at reap to record the usage rollup.
- pricing.py: add grok-build-0.1 rates ($1/1M input, $0.20 cached, $2/1M
output), verified against xAI's published pricing. Grok is a priced
non-Anthropic model, so cost computes the moment usage is captured.
- secret-scrub.js: an opencode tool.execute.before plugin porting the
security-critical bash-guard deny rules (git network ops, credential-file
reads, /proc env, internal-host HTTP, roboco.* imports, ROBOCO_AGENT_ID
forgery, env dumps, destructive rm) to the opencode runtime — restoring the
guard the Claude Code hook can't provide there. Throwing denies the call
(confirmed by opencode's env-protection example). Wired into the generated
opencode.json plugin array + baked into the grok image.
Deny logic verified via node (9 deny + 5 allow cases). UNVALIDATED against a
live opencode runtime: confirm it fires in the live E2E spawn before a Grok
dev-agent touches a real repo; the bash permission is operator-tunable as a
second gate.
Cost CAPTURE (distinct from pricing) is intentionally NOT built yet: opencode's
plugin hooks expose model info but no token/usage object, so the capture path
is unconfirmed and needs the live spawn to settle.
Complete the native Grok (xAI) path so grok-build-0.1 runs as a real
RoboCo agent, not just the provider seam.
- roboco-agent-grok image (docker/agent-grok.Dockerfile): FROM agent-base
+ opencode (the OpenAI-protocol runtime). One image serves every role;
role behaviour comes from the mounted manifest / mcp-config / system
prompt, exactly as on the Claude path.
- Entrypoint renders opencode.json at spawn from the GrokProvider env
contract + the mounted Claude Code mcp-config.json
(roboco.llm.providers.opencode_config): translates RoboCo's gateway
servers (roboco-flow / roboco-do / ...) into opencode's mcp block,
declares the xAI OpenAI-compatible provider + model, and wires
permissions + instructions. Pure, unit-tested translation.
- Orchestrator registers GrokProvider with the registry-qualified image
(_qualify_agent_image) so it resolves in local and registry deploys.
- Compose (both files + the registry compose) gain an agent-grok-image
builder service.
- Panel: a Grok (xAI) API key card on the AI Providers page, plus the
grok ModelProvider value.
KNOWN PARITY GAP (opencode runtime): the bash-guard PAT-scrub and the
transcript-based usage/cost capture are Claude Code hooks and do not
transfer to opencode. bash permission is operator-tunable
(ROBOCO_GROK_BASH_PERMISSION) so a deployment can fail closed until a
security/usage-parity opencode plugin lands. That plugin and live E2E
validation are the remaining work to finalize with xAI.
Add a roboco/llm/providers/ seam — an AgentProvider lifecycle ABC and a
ProviderRegistry keyed by ModelProvider — so the orchestrator can drive
agent backends other than Claude Code.
The first non-Claude backend is GrokProvider for xAI's grok-build-0.1.
xAI is OpenAI-compatible only (no Anthropic-Messages endpoint), so a Grok
agent runs an OpenAI-protocol runtime pointed at https://api.x.ai/v1
rather than the ANTHROPIC_BASE_URL injection the other providers use. It
reuses the orchestrator's existing mount/auth assembly, so it inherits the
same MCP gateway + tool-manifest wiring as every other agent by
construction, and passes its prompt via env (never an argv positional).
The change is purely additive: only GROK routes through the registry;
Anthropic / Ollama Cloud / self-hosted spawns run the existing
_spawn_container path unchanged.
Includes:
- ModelProvider.GROK (migration 038) + a seeded Grok provider row
(migration 039) + a grok-build-0.1 catalog entry
- GET/PUT /api/providers/grok-key to store the xAI key (Fernet-encrypted,
reusing the existing provider-key machinery)
- ClaudeCodeProvider reference adapter over the current spawn
- unit tests for the registry, GrokProvider (gateway wiring, no
ANTHROPIC_* leak, prompt-injection safety, failure paths) and routing
The dedicated roboco-agent-grok image and the exact OpenAI-protocol CLI
invocation are the remaining piece to finalise with xAI.
* [0c7a4732] feat(cockpit): add completed_30d and median_lead_time_hours to delivery summary (#207) (#210)
- Extend DeliverySummary schema with completed_30d: int = 0 and
median_lead_time_hours: float | None = None fields
- Add TaskService.get_delivery_stats_30d() that queries tasks completed
in the last 30 days and computes statistics.median of lead times
- Update CockpitService.summary() to source both new keys from
get_delivery_stats_30d() and include them in the delivery dict
- Update tests: mock new method in _patch(), assert new fields in
test_summary_aggregates, fix test_route_ok_for_ceo dict, add three
new unit tests for get_delivery_stats_30d (empty, multi, single)
Co-authored-by: Backend Developer 1 <be-dev-1@agents.roboco.dev>
* [d2647edf] Frontend: Build CompanyScorecard card on Goals tab (#211)
* [12569f37] Extend CockpitSummary type and build CompanyScorecardCard component (#208)
* [12569f37] feat(cockpit): extend CockpitSummary type with completed_30d and median_lead_time_hours
Add optional delivery.completed_30d (number) and top-level
median_lead_time_hours (number | null, optional) to CockpitSummary
interface in panel/src/lib/api/cockpit.ts so the API shape captures
the new backend fields without breaking existing consumers.
* [12569f37] feat(business): add CompanyScorecardCard component
Create panel/src/components/business/company-scorecard-card.tsx
exporting CompanyScorecardCard. The card fetches /cockpit/summary
via useQuery and renders five always-visible sections:
- Delivery: in_flight, blocked, awaiting_ceo, completed_30d tiles
(all from API response; no hardcoded numbers)
- Spend: 30d spend + projected monthly; muted 'No budget cap set'
when cap is null; red/destructive styling only when cap is a
non-null number AND over_budget is true
- Speed: 'X.Xh median — target: < 24h' when value present;
'No data yet' when null/undefined; '0h' never rendered
- Two stub Objectives with 'Not tracked yet' label, muted text,
and dashed-border styling — no fabricated numeric values
- Loading: three grouped Skeleton blocks
- Error: OfflineState with title 'Could not load scorecard data'
---------
Co-authored-by: Frontend Developer 1 <fe-dev-1@agents.roboco.dev>
* [f1f5cded] Integrate CompanyScorecardCard into GoalsTab and pass quality gate (#209)
* [f1f5cded] feat(cockpit): extend CockpitSummary with completed_30d and median_lead_time_hours
Add optional delivery.completed_30d (number) and top-level
median_lead_time_hours (number | null, optional) to CockpitSummary
interface in panel/src/lib/api/cockpit.ts. Backward compatible.
* [f1f5cded] feat(business): add CompanyScorecardCard component
Create panel/src/components/business/company-scorecard-card.tsx
exporting CompanyScorecardCard. Fetches /cockpit/summary via
useQuery and renders five always-visible sections: Delivery (no
hardcoded numbers), Spend (muted 'No budget cap set' when null;
red only when cap set AND over_budget true), Speed (X.Xh median
or 'No data yet'), two stub Objectives with dashed border and
'Not tracked yet' label. Loading: three skeleton groups. Error:
OfflineState 'Could not load scorecard data'.
* [f1f5cded] feat(goals-tab): integrate CompanyScorecardCard into GoalsTab
Import and render CompanyScorecardCard below the charter form in
goals-tab.tsx. The scorecard fetches its own data independently
so all loading/error states are handled per-card. Both cards are
always rendered in the Goals tab.
* [f1f5cded] fix(scorecard-tests): add vitest framework and CompanyScorecardCard test suite
Install vitest + @testing-library/react + @testing-library/jest-dom +
jsdom + @vitest/coverage-v8 as devDependencies in panel/.
Add panel/vitest.config.ts (jsdom env, @/* alias, coverage on
company-scorecard-card.tsx with 80% threshold).
Add panel/src/test/setup.ts (jest-dom matchers).
Update panel/package.json: add test, test:watch, typecheck scripts.
Update panel/eslint.config.mjs: ignore coverage/ directory to keep
lint clean of generated files.
Write panel/src/components/business/__tests__/company-scorecard-card.test.tsx
with 8 tests covering all 7 AC2 scenarios:
- loading skeleton rendered
- OfflineState on error
- OfflineState when data undefined
- delivery counts from mock data
- spend 'No budget cap set' when cap null
- spend destructive styling when cap non-null and over_budget true
- speed 'No data yet' when lead time null
- speed formatted value when lead time present
pnpm lint: 0 errors pnpm typecheck: 0 errors
pnpm test: 8/8 pass coverage: stmts 95% branches 90% fns 91% lines 95%
---------
Co-authored-by: Frontend Developer 1 <fe-dev-1@agents.roboco.dev>
---------
Co-authored-by: Frontend Developer 1 <fe-dev-1@agents.roboco.dev>
---------
Co-authored-by: Backend Developer 1 <be-dev-1@agents.roboco.dev>
Co-authored-by: Frontend Developer 1 <fe-dev-1@agents.roboco.dev>
The self-heal origination tests blind-inserted the fixed-uuid foundation
system agent (and a main-pm agent). In the full CI suite the app lifespan
seeds + commits those agents first, so the insert hit a duplicate-key on
pk_agents — green in isolation, red in CI. Get-or-create both (by id / by
slug) so the tests pass whether or not the agents already exist. Verified
against the real ordering (an app-lifespan integration test before this
file): 187 passed.
Two hardening fixes from the gap review:
- Optional self_heal_ci_workflow scopes the CI signal to one workflow file
(the workflow-scoped Actions endpoint). Without it, "latest completed run
across all workflows" could miss a red CI run masked by a later passing
workflow, or false-trigger on a non-CI workflow — unreliable on a
multi-workflow repo.
- The loop logs a warning when self-heal is armed but self_heal_project_slug
is unset, so a misconfiguration isn't mistaken for "all green".
Tests cover the workflow-scoped endpoint.
Adversarial review found the load-bearing invariant broken at the dispatch
layer: _dispatch_pm_work skipped only PR_REVIEW_SOURCES, so a PENDING
team=main_pm self_heal task (assigned_to=None, confirmed_by_human=False) was
routed to Main PM and spawned BEFORE the CEO approved it — the "never start
until you approve" promise didn't hold.
Fix: the PM dispatcher now also skips source='self_heal' while
confirmed_by_human is False (before the assigned/unassigned split, so it holds
either way); the task still shows in the panel so the CEO can see and approve
it. approve_and_start flips confirmed_by_human=True (the CEO's start IS the
human confirmation), so it dispatches normally afterward. Other sources are
unaffected.
Tests: a unit test that the dispatcher holds an unconfirmed self_heal task but
routes a confirmed one and ordinary tasks, plus a DB test that approve_and_start
flips the gate. (The readiness gate was deliberately not used — a blocker there
marks the task `blocked`; the dispatch skip leaves it cleanly PENDING.)
The CI telemetry call is the feature's only real-world I/O and was previously
exercised only through a fake source. Cover the GitHub Actions request shape
(/actions/runs, branch/status/per_page, auth) and response parsing, plus the
safe-None paths (missing token, GitHub error, no runs).
Register _self_heal_loop alongside the other background loops (created in
start, cancelled in stop). It returns immediately unless self_heal_enabled, so
a standard deployment adds zero behaviour and makes no CI call; when on it runs
one engine cycle per interval and commits any opened fix task. A test pins the
default-off dormancy (no sleep / CI / DB when disabled).
Behind the second opt-in (self_heal_originate_enabled), a detected regression
also opens a fix task into RoboCo's own delivery lifecycle and STOPS: PENDING,
unassigned, confirmed_by_human=False, team=main_pm, source=self_heal, with
synthesized acceptance criteria and a self_heal_fp= dedupe marker. It rides the
normal dispatchers only once the CEO Approve-&-Starts it; the loop itself never
calls start / approve / merge / deploy.
- TaskService: SELF_HEAL_SOURCE, extract_self_heal_fingerprint, and
list_open_self_heal_tasks (the dedupe + open-cap basis)
- SelfHealEngine._originate: per-signal fingerprint dedupe, per-cycle and
rolling open-task caps, repo resolved to RoboCo's own project (notify-only
when it can't be resolved)
- 7 DB-backed tests including the never-start / never-approve invariant
The detect side of the self-healing loop, modeled on the strategy engine: a
pure assess() turns breaching telemetry samples into RegressionObservations
(with a stable per-signal fingerprint for later dedupe), and run_cycle() is a
no-op unless self_heal_enabled and otherwise only sends the CEO one
ack-notification per regression. Detect + notify only — it never originates,
starts, merges, or deploys; the telemetry source is injectable for testing.
6 unit tests.
First slice of the production self-healing loop: a read-only telemetry source
that watches RoboCo's OWN repo CI and normalizes the latest GitHub Actions run
conclusion into breach / no-breach samples for the regression detector. It
targets only the single project named by self_heal_project_slug — RoboCo
healing itself, never other/client repos; the org's repo-agnostic delivery flow
is untouched.
- config: self_heal_enabled / self_heal_project_slug / self_heal_originate_enabled
plus interval and open-task / per-cycle caps, all default-off
- GitService.get_latest_ci_conclusion: per-project Actions-run lookup (graceful
None on missing token / no runs / error; never raises into the loop)
- TelemetrySample + TelemetrySource contract + GitHubCITelemetrySource
- 5 unit tests
Real-Postgres round-trips for the external/internal PR-review TaskService
helpers that the existing mock tests can't prove actually persist:
- ingest_external_pr — create-once, head-SHA dedup (unchanged head skips),
and re-review on a new head; internal_pr source wording;
- pr_review_claim / complete_review — the planless, branchless
pending -> in_progress -> completed lifecycle, with re-claim / re-complete
no-ops and the "complete requires in_progress" guard;
- create_supersede_umbrella / find_supersede_umbrella — created on the same
repo (not parented), idempotent lookup, non-review rejection, and the
pr=5-vs-pr=50 exact-marker disambiguation;
- list_external_pr_reviews — source isolation, the data-layer half of the
dispatcher contract (regular tasks never leak into the review queue).
Writes use flush (not commit) so the rollback-per-test fixture keeps each
case isolated from the others against the shared session-scoped test DB.
Add a Feature Flags card to the Settings page that toggles env-gated
subsystems (external/internal PR review, web research, strategy engine,
pitch provisioning, RAG auto-update, transcript pruning) directly from
the panel instead of hand-editing environment variables.
Flags persist in system_settings as 'true'/'false' and are overlaid onto
the live config singleton at startup; an unset flag keeps its
environment/config default. A toggle takes effect on the next backend
restart — no per-consumer re-routing.
Backend: FEATURE_FLAGS registry + bool validator + get_bool accessor on
SettingsService; feature_flag_effective_values and
apply_persisted_feature_flags; GET /settings/feature-flags; best-effort
startup overlay in the app lifespan.
Frontend: settingsApi.getFeatureFlags / setFeatureFlag and a
FeatureFlagsCard rendered full-width below the settings grid.
A coordination root (team=main_pm, product-linked, no repo) the CEO sends back
lands in needs_revision, but the dev dispatcher skips it (not a cell team) and
the closure path only handles paused parents — so it sat in needs_revision
forever. (NOT a foundation-spec gap: the spec already allows needs_revision ->
claimed for any role.)
- _dispatch_revision_coordination_roots: re-spawn the owning PM for a
needs_revision coordination root so it re-coordinates the revision (registered
in the dispatch loop after PM closure)
- _readiness_check_role_for_status: widen the dev-owned states (needs_revision,
verifying) to also accept cell_pm/main_pm for coordination roots — a pure
widening; normal code tasks stay dev/doc-only
- 16 unit tests (dispatcher decision + readiness widening)
Covers the inbound-PR read surface: list_open_prs normalization + fork/internal
classification (and the recent _fetch_open_prs/_normalize_open_pr refactor),
plus get_pr_diff's diff-media-type request — both with their safe-empty paths
on missing token / GitHub error. The DB-backed paths
(ingest/complete_review/pr_review_claim/supersede umbrella) are covered
separately.
Renn F