The review question sentence was written in four places and the URL
safety rules in three. Collapse each to one definition.
- Add `ReviewTarget::question_text_with_link` as the single definition of
the question wording. `question_text()` and the Slack header both use
it, so a wording change is now one edit.
- Delete `ReviewTargetKind::noun()`. The enum already derives
`strum::Display` with the same snake_case output.
- Share one `review_target_line` helper between the console interviewer
and the CLI attach client, which held a byte-identical copy. Print only
the URL: `question.text` already carries the label and the noun.
- Trim the web-side check to the URL scheme, host, and credentials, which
are what a raw `href` can act on. Label length and control characters
cannot affect the DOM and stay server-side.
- Split validation from presentation in the web UI. `safeReviewTarget`
returns the target or null, and each caller picks its own fallback, so
an unsafe target now falls back to the same Markdown rendering as a
question with no target.
- Derive the resource noun from `kind` in the web UI instead of
hardcoding "document".
- Use `ReviewTargetKind.DOCUMENT` and the shared `isRecord` guard when
parsing events, instead of a raw string and a hand-rolled object check
that accepted arrays.
- Drop `deny_unknown_fields` from the wire struct. The OpenAPI schema
leaves `additionalProperties` permissive, so an added field would
otherwise make persisted events unreadable.
- Import `ReviewTarget` by name, and stop naming Slack in a fabro-types
error message.
- Document that `review_target=true` replaces the gate's `label`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Follow-up cleanup on the publish-failures change.
Error model:
- Collapse `Error::{Engine, Publish, Handler}` into one `Error::Stage` with an
`ErrorStage` discriminator. The three shared a field shape and had to be
edited together in four match groups; nine near-identical constructors
become two private helpers.
- Add `Error::failure_reason()`, replacing the same error -> FailureReason
mapping written out in four places.
- Publish errors are now terminal. Publish runs once, after execution, so no
caller could ever act on the retryable classification.
Publish phase:
- Fix: a branch that was pushed is now still reported when pull request
creation fails afterwards. `PublishOutcome` records what happened and
carries the error separately, instead of hiding both behind a `Result`.
- Drop `PublishOutcome::NoChanges`, which no consumer distinguished from
`Published { pr_url: None }`.
- Move publish onto `Concluded` as methods and replace three near-identical
precondition guards with one `publish_target()`.
Pull requests:
- `maybe_open_pull_request` -> `open_pull_request` returning the record
directly. Both callers already reject empty diffs, so the `Ok(None)` path
was unreachable.
- Drop `CreatedPullRequest.head_sha`, which echoed back its own input.
GitHub client:
- Delete `branch_exists`, which had no callers and duplicated
`branch_head_sha`. Give `branch_head_sha` the `_with_client` split every
sibling has and port the tests to `MockHttpClient`.
- Collapse the copy-pasted credential match in `resolve_clone_credentials`.
Events:
- `PullRequestCreated.head_sha` is `Option<String>` instead of using an empty
string to mean absent.
- Centralize the run-branch refspec in `lifecycle::push_run_branch`, so
`git.push` reports a branch name from both emitters as documented.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
- Take `Arc<Catalog>` by value again through the validation entry points.
`AppState::catalog()` returns an owned `Arc`, so `&state.catalog()` was
cloning, borrowing the temporary, then cloning again at the leaf. Every
consumer ends up owning the `Arc`, so by-value is the honest shape and it
drops one clone per call. The one caller holding the catalog in a field
now says `Arc::clone(&self.catalog)` explicitly.
- Correct the `RenderMode` doc comment. It claimed `Strict` is "used by
run-create", but run-create renders `Structural` and promotes the
resulting warnings to errors itself; `Strict` has no production caller.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`ModelResolutionOptions` was a field-for-field duplicate of the existing
public `ModelResolutionTransform`, down to a verbatim copy of its `new()`.
`pipeline::transform` then unpacked one to rebuild the other, cloning the
catalog Arc and the eligible-provider set on the way.
- Delete `ModelResolutionOptions`. `TransformOptions.model_resolution` now
holds an `Option<ModelResolutionTransform>` directly, so the TRANSFORM
step is `resolution.apply(graph)?` with no rebuild and no clones. This
is consistent with `custom_transforms`, which already holds transforms.
- Add `ModelResolutionTransform::catalog()` so the VALIDATE step can reach
the same catalog for its lint rules. That is the only new code needed.
- Drop `CatalogScope` from `operations::validate`, which was a third copy
of the same fields. The three entry points now hand a partially built
transform to `validate_resolving_models`, which completes it with the
workflow's default provider once the workflow is resolved.
- Extract `validate_child_workflow` in `manager_loop`, collapsing two
near-identical validate-and-unwrap blocks.
- Point the transform tests at their own `transform_options()` helper via
struct-update syntax instead of respelling all seven fields, and drop a
HashSet -> Vec -> HashSet round trip from the create test helper.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Follow-up cleanup on the catalog-free validation split. Same behavior,
fewer parallel code paths.
- Make the catalog an explicit `Option<&Catalog>` on `pipeline::validate`
instead of a `validate` / `validate_with_catalog` pair, so each call
site states whether catalog rules run.
- Collapse `preprocess_and_validate`, `preprocess_and_validate_structural`,
and `preprocess` into one function that takes `TransformOptions`. Its
`model_resolution` field is now the single source of truth for catalog
awareness, which drops a 12-argument signature and the
`too_many_arguments` allow.
- Replace the duplicated resolve-and-preprocess block in
`operations::validate` with one `validate_in_scope` helper, and drop the
HashSet -> Vec -> HashSet round trip on the catalog path.
- Extract `configured_default_provider`, previously duplicated between
`operations::create` and `operations::validate`.
- Delete `validate_manifest_with_environment_defaults`, which had no
callers outside its own module.
- Share the `server-model.fabro` fixture between the two CLI tests instead
of inlining it twice. The validate test now asserts the rendered output
through the usual snapshot helper, which also removes a hand-rolled
`std::fs::write` and its clippy allow.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A Daytona-backed run could fail five seconds after start when the sandbox
git clone hit a transient GitHub "Repository not found" error. Clone-based
providers mint an installation access token and clone with it in the same
breath, but GitHub replicates a new token to its edge cache sites
asynchronously. A clone that starts within a second of the mint can be
rejected before the token is visible to the site serving it, and on a
private repo that rejection arrives as "Repository not found" because
GitHub answers unauthorized reads with 404.
Nothing retried the clone, and the failure classified as `deterministic`,
which is the one category `loop_restart` refuses to restart. An identical
run relaunched 46 seconds later succeeded with no changes.
A successful mint is what makes the message safe to retry.
`resolve_clone_credentials` already fails loudly on every deterministic
explanation for a clone 404: the installation lookup 404s when the App is
not installed for the owner, and token creation 422s when the installation
does not cover the repo. Once credentials are in hand, "not found" from the
clone itself cannot mean "no access".
Add `clone_retry` and use it from both clone-based providers: 3 attempts
with 3s then 9s backoff, reusing the same token so replication keeps making
progress instead of restarting the clock. Token-replication signatures
retry only when credentials are present, so a public clone of a wrong URL
still fails fast. Infrastructure failures retry either way.
The Docker provider had the identical single-shot clone and is the default
runtime provider, so it is covered too.
Also fix two nearby issues found while reading the area:
- The GitHub-URL-parse path in the Daytona clone skipped `fail_init`,
unlike every sibling path, so `InitializeFailed` was never emitted.
- The `classify_exec_failure` hint for "repository not found" asserted the
App installation may not cover the repo. After a successful scoped mint
that diagnosis is impossible, and it sent operators hunting a
configuration problem that did not exist.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The process environment is no longer a configuration source. `{{ vars.NAME }}`
(non-sensitive, server-stored) and `{{ secrets.NAME }}` (vault-backed) cover
both cases, and reading the worker's ambient environment made a run's inputs
depend on how its process happened to be launched.
`Namespace::Env` is kept but wired to nothing, so `{{ env.NAME }}` still
parses and fails with a message naming its replacement rather than reaching
a consumer as literal text. `ResolveCtx::with_env` is gone, so no call site
can opt back in.
Two long-standing warts were env-only and go with it:
- `InterpString::resolve_or_source`, the "fall back to the raw template
source on failure" path, which let an unresolved token reach a sandbox or
the GitHub API as literal `{{ ... }}` text. Its own comment noted it was
slated for hard-error semantics.
- `RunEnvironmentSettings::resolve_env`'s matching source fallback for
env-only values.
Both carried `#[expect(clippy::disallowed_methods)]` escape hatches. Every
run-boundary resolver — sandbox env, prepare steps, MCP transports, GitHub
permissions, Slack channels, run goal files, provider extra_headers — now
fails closed instead.
Hooks lose their `allowed_env_vars` allowlist, `resolve_header`, and
`HeaderResolveError` along with the `E: Env` generic threaded through the
executor. They keep `{{ vars.* }}`, which `RunSettings::substitute_variables`
already substitutes server-side at run creation.
`allowed_env_vars` is removed from the OpenAPI spec and the generated
TypeScript client. The docs example showing `{{ env.* }}` in
`[server.slatedb.s3].bucket` was already wrong — that field is a plain
String and never interpolated — and is now a literal.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`EnvCredentialSource` resolved provider credentials from the process
environment. It had no production entry point of its own — it was only
ever reached as the `None` arm of an `Option<Vault>` in three places:
`build_llm_source`, `configured_providers_for_start`, and
`configured_providers_from_process_env`.
That optional vault is not a state the product can be in. Every run has a
server behind it, the server always spawns workers with `--storage-dir`
(`worker_runtime.rs`), and `SqlVaultCredentialSource` backs both the
server and the CLI. So the fallback only served to silently degrade
credential resolution to whatever the worker process happened to have in
its environment.
Make the vault required across the run path — `RunOptions`,
`StartServices`, `build_llm_source`, `tool_secrets_from_configured_sources`,
`vault_token_lookup`, and the CLI GitHub helpers — so the invariant is
enforced by types rather than assumed. A worker spawned without
`--storage-dir` now fails with a clear message instead of quietly
continuing without a vault.
`configured_providers_from_process_env` had no callers at all and is
deleted. `AgentApiBackend::new_from_env` was public but only ever called
from its own tests; it is deleted too.
Test-only credential sources move to a feature-gated
`fabro_auth::test_support`, wired through dev-dependencies so they never
link into production builds. The CLI worker tests now pass
`--storage-dir`, matching what the server actually does.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Command node `script` attributes were literal text: a `{{ inputs.x }}`
reached bash verbatim, and the only signal was a `detemplated_attribute`
warning. Scripts now substitute `{{ goal }}`, `{{ inputs.NAME }}`, and
`{{ vars.NAME }}` at run creation, alongside goals and prompts.
Scripts use `InterpString` token substitution rather than the MiniJinja
pass that renders prompts. Shell source is full of brace syntax that must
survive untouched — jq filters, awk programs, Go templates, brace
expansion — and `InterpString` claims only the narrow token forms,
leaving everything else literal.
`env` and `secrets` are deliberately not wired and now fail loudly
instead of passing through as text. A script reads the environment with
`$NAME`, which needs no interpolation, and a resolved secret would be
baked into the `CommandStarted` event that records the script verbatim.
The error points at `[environments.<slug>.env]` for the secret case.
`ResolveCtx` gains opt-in `with_inputs` and `with_goal`. Namespace
availability stays scope-determined per call site, so every existing
config-layer context leaves both unwired and keeps its current behavior.
`goal` names a single value rather than a namespace of them, so it has
no dotted form: only the exact body `goal` produces a token and
`{{ goal.title }}` stays literal.
Values substitute verbatim without shell quoting, matching
`[[run.prepare.steps]].script` where the snippet is the author's to
quote. Substituted text is never rescanned.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The model indicator on a stage page hovered to provider, model, and
reasoning effort only. Seeing what a stage actually spent meant leaving
for the Billing tab, which reports per node rather than per visit.
The stage list had no token data to show, so add a per-visit `billing`
block to `GET /runs/{id}/stages`. The Billing tab's pricing rule (a
provider-reported cost wins, otherwise the server catalog prices the
tokens) was private to `billing_rollup`; move it to
`StageProjection::billed_usage` and drive both call sites from it so the
two views cannot drift.
The popover's buckets use the Billing tab's labels verbatim. It stays
scoped to one visit, so a looped node's row on the Billing tab is the sum
of what each of its visits shows here.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A node with no `shape` defaulted to `box`, which resolves to the agent
handler. That made a shapeless `script` node run as an LLM call prompted
with its own label, while the `script` was reported as inert — wrong
behavior behind a warning.
`script` is read by the command handler and by nothing else, so a
shapeless node that sets it is unambiguously a command node. `shape()`
now infers `parallelogram` in that case. An explicit `shape` still wins.
Two rules keep the inference honest:
- `script_prompt_conflict` — setting both `script` and `prompt` is an
error. No handler reads both. It fires regardless of shape so that
adding one cannot downgrade the error to a warning.
- `command_requires_script` — a command node without a script is an
error. Without this the original trap just moves: a node meant as a
command that omits its script silently becomes an agent again.
Also drops the `tool_command` alias in favor of `script` alone, routing
the six read sites through a new `Node::script()` accessor.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The DOT parser created a node for every edge endpoint, and nothing
recorded whether a node came from a declaration or was synthesized from
an edge. The edge_target_exists rule only checked whether the node id
was present in the graph, which was always true by then, so a misspelled
endpoint became an attribute-free node that defaulted to shape=box — an
LLM stage. Validation emitted a prompt_on_llm_nodes warning and exited 0.
Node now carries `implicit`, set only when the parser synthesizes the
node from an edge endpoint. A declaration anywhere in the workflow
clears it, so order does not matter and subgraph declarations count.
Node::new leaves it false, so programmatic construction and graphs
deserialized from older checkpoints read as declared.
edge_target_exists treats an endpoint as valid only when it exists and
is declared, reporting each undeclared node once. The near-identical
missing-source and missing-target branches collapse into one path. The
import transform copies the flag onto spliced nodes so an edge-only node
inside an imported fragment is caught too.
parse_and_validate_human_gate had two edge-only nodes and now declares
them; it was an instance of the bug rather than a casualty of the fix.
No shipped workflow, docs example, or CLI fixture relied on the old
behavior.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`Claude5QuestionToolArgs`/`Claude5Question`/`Claude5Option` differed from
the Anthropic trio only in required-ness -- `header: String` rather than
`Option<String>`, same for each option's `description`. The JSON Schema
already enforces that at the model boundary, so the lenient structs
deserialize the strict payload unchanged.
`normalize_claude5_questions` then reproduced `normalize_anthropic_questions`
plus an inlined copy of `options_from_anthropic`, so `option_key`,
`display_text`, and `bounded_display_field` were each applied in two
places and could drift.
Replace both with one normalizer taking a `QuestionLimits`. The genuine
Claude 5 deltas -- at most four questions, two to four options, a
twelve-character header cap, required header and option descriptions, and
no previews on multi-select -- become data rather than a second code path.
Two rules serde used to enforce are now the normalizer's: a missing header
and a missing option description. Both are still rejected, with a clearer
message than serde's "missing field". `multiSelect` now defaults to false
instead of being a deserialization error; the schema still marks it
required, which is where that contract belongs.
Adds tests pinning the strict rules against the shared normalizer, and one
asserting the lenient contract still accepts optional headers and
descriptions.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
All six provider profiles embed a `BaseProfile` and hand-wrote the same
six delegating accessors -- 24 identical lines each. What actually
distinguishes them is `build_system_prompt`, and for Claude 5,
`register_subagent_tools`.
Replace the copies with one `impl_base_profile_accessors!()` invocation.
A macro rather than trait defaults because three implementors have no
`BaseProfile` to delegate to -- `TestProfile`, the workflow crate's
`ShutdownTestProfile`, and the server's `AskFabroProfile` -- so a default
would need a runtime fallback for a case the compiler can already rule
out. Those three keep their hand-written accessors and are untouched.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`NativeTool` documents itself as "an identity, not a name" whose canonical
form is fabro's own vocabulary, with harness names layered on as aliases:
`to_string = "read_file", serialize = "Read"`.
The four Claude 5 subagent tools inverted that. `ClaudeAgent` declared
`to_string = "Agent"`, making the Anthropic wire name the identity and
leaving `name(ToolVocabulary::Fabro)` returning `"Agent"` -- and pairing a
provider-specific variant name with a generic wire name. It also meant the
`Claude5` arm listed none of them: they fell through to
`canonical_name()` and were correct only by accident.
Rename to `BackgroundAgent` / `AgentOutput` / `StopAgent` / `MessageAgent`
with fabro canonical names, keep the harness names as `serialize` aliases
so `from_any_name` still resolves them, and name them explicitly in the
`Claude5` vocabulary arm. Also map `Grep`/`Glob` there: that arm describes
the vocabulary rather than the profile's registry, and if either were ever
registered it would otherwise reach the harness lowercased.
Records why these are separate identities from
`spawn_agent`/`wait`/`close_agent`/`send_input` rather than aliases of
them, since the capabilities genuinely differ.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Task tools scope their list by `root_session_id` -- `Session` documents
this as "a subagent session inherits its parent's `root_session_id` so
todo tools that scope by root (Anthropic tasks) share one list across all
subagents" -- so a root and its children address one logical list.
`build()` runs once per session, though, and `AnthropicProfile`
constructed its own `TodoRuntime` inside that call. Root and child
therefore resolved the same `list_id` through different runtimes: both ID
counters started at zero, so both emitted `todo.created` with id `1` for
the same list, and `TodoListProjection::upsert` matches on id -- the
child's task replaced the parent's in the persisted projection. `TaskGet`
and `TaskList` read the local runtime, so neither session could see the
other's tasks either.
The previous commit's shared runtime fixed this for Claude 5 only,
because `build()` passed dependencies positionally and adding a fourth
argument would have meant touching all six call sites. It grew a second
constructor for Claude 5 instead, leaving the other five on a signature
that could not carry the runtime.
Bundle them into `ProfileDeps` so every profile takes the same
`(model, &deps)`. The duplicate constructor is gone, Anthropic shares the
runtime by construction rather than by opting in, and a future dependency
reaches all six profiles or none.
The existing Claude 5 sharing test is generalized and now also runs for
Anthropic; it fails against a per-profile runtime.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Follows the type rename: a rotation chain is now an auth session with its own
row, so the local names and the Repository doc comment should say so rather
than referring to a store that no longer exists.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Removes `slate/auth_tokens.rs`, `Database::refresh_tokens()`, and its
`OnceCell` now that nothing reads them, and clears the retired `auth/refresh`
prefix once at startup. That sweep is not housekeeping we could skip: the
reaper that used to collect those records went with the store, so without it
they would sit in the object store forever. A later boot finds the prefix
empty and does nothing.
`record/transaction.rs` goes too -- rotation was its only caller, and SQLite
transactions replaced it. `KeyedMutex` stays; `AuthCodeStore` still uses it
until auth codes move.
Existing refresh tokens are not migrated. Everyone re-authenticates once on
upgrade.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`ParentNotificationHub` kept a second `Mutex` and `watch` channel holding
a copy of each child's terminal result -- data `SubAgent.status` already
owns as `SubAgentStatus::Finished`, and which is never evicted, since
nothing removes entries from `SupervisorState.agents`.
Two of the three bugs fixed in the previous commit were ordering bugs in
the coupling between those two structures: suppress-vs-commit in
`begin_shutdown`, and register-vs-publish in `spawn_inner`. Both were
fixed by ordering the steps correctly. Keeping the registration beside
the status it is delivered with makes that whole class unrepresentable
instead:
- Registration is now a field on the `SubAgent` literal `spawn_inner`
already builds, under the lock that publishes it. There is no window
between publishing an agent and registering its notification.
- Suppression on shutdown happens inside the critical section that
decides the shutdown, after the status transition commits, so a
rejected shutdown cannot discard a result the parent is owed.
- `next_parent_notification_batch` scans agents for a live registration
whose status is `Finished`, and ignores `Closing`/`Closed` outright --
so a shutdown racing delivery can no longer park the parent on a result
that will never arrive, even if suppression were missed.
`spawn_result_monitor` no longer takes the hub; it bumps a single
`watch` counter after committing the status it already commits. Batch
order was the queue's insertion order, so `SubAgent` carries a
`spawn_seq` to keep delivery oldest-first.
Tests move from exercising the hub directly to the supervisor API, and
cover spawn-order batching and the shutdown-races-delivery case that the
old shape could not express.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every operation the SlateDB store answers with a full keyspace scan becomes
an indexed query here: listing a user's sessions joins one row per session
via the partial unique index instead of scanning every token ever issued
and grouping by chain, and revoking one is a single DELETE that cascades.
Rotation is the structural win. Claiming the presented token is one
`UPDATE ... WHERE used_at_ms IS NULL ... RETURNING`, and it is the
transaction's first statement, so SQLite takes the write lock before
anything is read. A concurrent caller blocks on that lock and then sees the
token already spent, which is exactly the replay signal -- so the store
needs no `KeyedMutex` to serialise rotation, and the guarantee survives more
than one server process.
Expiry is checked ahead of reuse on the cold path, preserving the ordering
callers depend on: only replaying a still-live token revokes its chain.
Drops the ordering CHECKs between a session's timestamps and its tokens'.
Rotation stamps `now` from the process clock against rows written by an
earlier request, so an NTP step backwards would have turned a harmless clock
anomaly into refresh failing outright for every affected session.
The store is not wired into the server yet.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Three correctness fixes in the Claude 5 background-agent path, plus
cleanups from a reuse/quality/efficiency review pass.
Fixes:
- Background-agent output was run through skill expansion. A child that
wrote a bare path ("cleaned up /tmp") failed the whole parent turn with
`Unknown skill: /tmp`, and a child whose output happened to name a real
skill had its report replaced by that skill's template. Synthesized
harness turns now skip expansion; only text the user typed can invoke a
skill.
- `begin_shutdown` suppressed the pending notification before deciding
whether a shutdown would happen. Stopping an agent that had just
finished rejected the stop *and* discarded the result the parent was
owed. Suppression now happens only once shutdown is committed.
- `spawn_inner` registered the notification after publishing the agent in
`state.agents`, so a concurrent `shutdown_all` in that window left a
pending entry the monitor never completes, and the parent's drain loop
would never see the queue as drained. Registration now precedes
publication.
- `TaskOutput.timeout` was declared `number` but parsed with `as_u64`, so
a schema-valid `30000.0` failed at runtime.
- Update the fabro-server alias test for the `sonnet` alias moving to
Claude Sonnet 5.
Cleanups:
- The supervisor renders the notification turn; `Session` no longer knows
the envelope format.
- Replace six near-identical prompt snapshots with a property test over
all eight conditional combinations, keeping the default and
all-conditionals snapshots for wording.
- Collapse `TodoRuntime`'s two mutexes into one.
- Read the prompt vocabulary from the registry instead of hardcoding it.
- Drop internal vocabulary from the `SendMessage` tool description.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The check tested for `never reconstruct it from memory`, a phrase the
Kimi edit description no longer contains after it was reworded to match
Kimi Code. A one-sided `!contains` against a literal cannot tell "the
phrase is absent because nothing leaked" from "the phrase is absent
everywhere", so it silently stopped protecting anything.
Assert the marker is present in Kimi's own description and absent from
the stock one. Removing the marker from the description now fails the
test instead of quietly disarming it, verified by doing exactly that.
Also correct the grep docs: all three sandbox implementations probe for
`rg` and fall back to POSIX `grep`, so the page should not imply a
single engine. Pre-existing, adjacent to the lines this branch touched.
Reported by Copilot review on #646.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Neither harness puts read-before-edit mechanics in the system prompt.
Kimi Code's `system.md` has no such section; the rules live in
`edit.md`, `write.md`, and `read.md`. Codex's prompts say nothing about
reading before an edit at all, and its editing guidance is attached to
`apply_patch`.
Drop the `# Reading Before Writing` section from the Kimi prompt and
carry its content in the Edit, Write, and Read descriptions, worded as
Kimi Code words it. Nothing is lost: every bullet in the removed
section was already covered by a tool description.
Two behaviors change to match upstream. Edit now says not to issue
consecutive edits against the same file, since the first invalidates
the second's `old_string` -- Kimi Code's stated reason. Read now says
not to re-read solely to confirm a write landed, which both harnesses
call out as waste; the previous prompt asked for exactly that re-read.
The gpt56 profile already followed the Codex split and is unchanged.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`active_time_ms` was only ever computed from terminal stage events, so a
stage still running contributed zero to the run rollup. A run parked in one
long agent stage reported 2m 8s of active time against 16m 53s of wall
clock — the two finished stages — while the running stage had been doing
continuous inference and tool work for over 14 minutes.
`live_run_timing` summed `filter_map(|stage| stage.timing)`, and
`stage.timing` is only written at finalization. Wall time ticked live off
`start_time`; active time did not tick at all.
Stage projections now accumulate brackets from the event log:
- Closing an inference bracket folds its span into `live_inference_ms`
instead of discarding it, including across retries, matching the
in-process stopwatch.
- Tool calls open a batch on the first outstanding call and close it when
the last one drains, so tools running concurrently within a turn count
once — the same span `execute_tool_calls` is bracketed by. Summing
per-call durations would over-count parallel tool use. Subagent tool
events are excluded; they run inside the root call's span already.
- `StageProjection::live_timing(now)` composes accumulators with any open
bracket, per handler: agent stages use the brackets, prompt and command
stages count elapsed time as inference and tool respectively, and
handlers that wait on a human, timer, condition, or child branches
report zero.
Active is clamped to wall per stage. A worker killed mid-turn leaves its
bracket open forever, and without the clamp it would tick up unbounded.
The clamp does not need to detect the dead worker: a stage cannot have been
active longer than it has existed. `watchdog.timeout` remains the authority
on whether a run is stuck. The clamp is deliberately not applied at run
level, where concurrent branches can legitimately sum past run wall time.
Timing is derived from events rather than emitted by the worker, so this
needs no event-schema change and applies to runs already stored.
`StageProjection.timing` keeps its terminal-only meaning, and the
authoritative breakdown still replaces the live estimate at terminal
events.
The billing endpoint had the same hole behind its `wall_only` fallback:
running stages reported zero inference/tool/active. Not visible in the
product, which renders only `wall_time_ms`, but wrong for any other
consumer of `GET /runs/{id}/billing`.
Parallel branch stages lose their breakdown permanently, even after
completion, because `parallel.branch.completed` carries only `duration_ms`.
That is a separate data-loss bug, tracked in #644.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`ReadBeforeWriteSandbox` blocked writes to any existing file the agent
had not read, tracked by a session read set populated only by
`read_file`, `grep`, `read_many_files`, and the Kimi `Read`.
The gpt56 profile has none of those. It mirrors Codex's tool contract --
`shell_command`, `apply_patch`/`edit_file`, `update_plan`, `web_search`
-- and reads through the shell, so its read set stayed permanently
empty and every edit to an existing file failed. In run
01KYD4360GN6SED4BYEVGYP4XT all 28 `edit_file` calls failed, 25 of them
on the guard. The agent read `package.json` with `sed` and `cat`,
hex-dumped it trying to diagnose the rejections, then routed around the
guard with `sed -i`, which the guard never covered. It prevented no
blind write; it converted content-anchored edits into an unreviewed
in-place shell rewrite.
Neither Codex nor Kimi Code enforces read-before-write at runtime.
Codex's `apply_patch` `Add File` overwrites an existing path silently;
Kimi Code's `Write` has no check at all. Both rely on the exact-match
requirement in their edit tools, which is stronger proof of inspection
than a read set, plus per-write approval.
Tool descriptions and the Kimi prompt keep telling the model to read
before editing -- that guidance matches Kimi Code's own `edit.md` and
still prevents `old_string not found` -- but no longer claim the
workspace refuses unread writes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`Automation::to_toml_string` and the `to_persisted` helper it wrapped
have had no production callers since automations moved from
`<storage>/automations/*.toml` into SQLite. Writes now serialize through
`canonical_bytes` for revision hashing; nothing renders an `Automation`
back to a TOML document.
Repoint the canonicalization test at `parse_persisted` + `canonical_bytes`
so it exercises the production path that actually produces the bytes the
revision hash is computed over.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
