The freestyle analyst can now break a large goal with dependency seams or
independent review/handoff points into a parent epic + DEPENDS_ON-linked children
in a single T2 approval, instead of N separate task_create calls. Parent
DEPENDS_ON every child (completes last); each child IMPLEMENTS parent. Resolves
depends_on by ref or index; rejects cycles, unresolved refs, missing title/goal,
and empty batches; same batch dedup + force_reason convention as task_create.
A session works one active task, so multi-task work is multi-session by
construction: the analyst names the single ready task this run works, the architect
threads only that one, and siblings are claimed by later runs via task_ready
(claim-driven; no scheduler, /tasks/next stays rejected).
Doctrine: analyst_freestyle.md picks one-task-vs-decompose and names the ready
task; architect_freestyle.md threads only that one; plus the L0 policy line.
freestyle_planning.toml analyst gains task_decompose (pinned by
FreestylePlanningWorkflowTest).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Bypassing a gate with force=true now requires a force_reason, recorded as a
"[force] ..." note on the task so the override is auditable (visible in history
and the context bundle) rather than a silent escape hatch. Applies to the agent
tools: task_create (dedup override) and task_update (claim over unmet blockers,
complete with no in-scope writes). force without a reason is rejected.
The REST POST /tasks create override (operator/CLI path) still takes a bare force
and is left for a separate call.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Block writing a file whose on-disk content changed since the session read it — a
concurrent/external edit the agent's view doesn't reflect. file_read records a
content hash on whole-file reads; the orchestrator's completion path now carries
the tool's structuredOutput (it previously dropped it, unlike SandboxedToolExecutor
— the two paths were inconsistent), so SessionContext.readHashes folds it.
StaleWriteRule compares the current hash (WorldProbe.contentHash) against the
read-time hash and BLOCKs a mismatch, telling the agent to re-read first. Files the
session wrote itself are excluded, so its own edits never look stale; partial reads
set no baseline.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Completing a task that declared affected_paths but saw no matching write this
session is now blocked ("marked done without doing it"), escapable with force.
TaskUpdateTool reads the session's writes through a SessionWrites port whose
adapter folds SessionContext (writes = recorded receipt.affectedEntities), and
matches them against the task's affected_paths globs.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claiming a task now emits SessionWorkingTaskEvent{taskId, affectedPaths} into the
session's own stream (via a SessionFactRecorder port + event-store adapter), and
SessionContextProjection folds it into SessionContext.activeTask. So a gate learns
"which task is this session working, and its scope" by reading the session stream
— no cross-stream scan, no kernel->tasks dependency. Re-emitted on an affected_paths
edit by the claimant, so the snapshot scope stays current; latest wins on replay.
This is the data source the write-scope gate needs (next).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Harden the claim warning into a block: task_update action=claim on a task with
unmet blockers is now rejected (no mutation) with the blocker list, instead of
claiming-with-a-warning. Escapable with force=true, which claims anyway and keeps
the "claimed despite unmet blockers" warning for the audit trail.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The doctrine says "search before creating", but that leans on the LLM; this is
the backstop against an agent re-creating a task across runs. TaskService.
findDuplicates matches active (non-terminal) same-title tasks, case- and
whitespace-insensitively — a DONE/CANCELLED look-alike is not a duplicate, since
that work may legitimately recur.
- task_create rejects a duplicate, naming the look-alike and offering force=true.
- POST /tasks → 409 with the duplicates, or force:true to override;
correx task create --force.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Bring the work graph into the agent loop:
- task_ready (T1): lists tasks ready to work now (TODO, all dependencies
satisfied) so an agent looking for work can pick one and claim it. Surfaces
work; never assigns (honouring the rejected /tasks/next). Registered in
TaskTools — left unwired in the request-driven example workflows, where it
would be noise; it belongs in an autonomous work-pull workflow.
- claim guard: task_update action=claim appends a WARNING naming any unmet
blockers, so an agent knows when it has jumped a dependency. Surfaced, not
hard-blocked — status is derived, so we don't fight the reducer.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Nothing asserted that the example workflows actually grant the task tools, so a
future edit could silently drop one and stay green. Pin allowed_tools per stage for
role_pipeline (analyst/implementer/reviewer), freestyle_planning (analyst), and the
shipped review_loop (implement/review). The existing review_loop test loaded an inline,
now-stale toml rather than the shipped file, so add one that loads the real file.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A freestyle stage's tools are LLM-authored in the execution_plan. At run time an
unresolvable tool name is silently dropped (resolve -> null -> mapNotNull), so the
stage runs toolless — e.g. a misspelled task_update means the implementation stage
quietly loses task tracking, invisible without a live run.
ExecutionPlanCompiler now takes the registered-tool universe and rejects a plan that
names an unknown tool, with a message identifying the tool and stage. FreestyleDriver
already turns a compile failure into a rejection the architect retries. The param
defaults to empty (validation skipped) so existing callers are unaffected; Main feeds
it toolRegistry.all().
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The task tools were registered but had only mechanical descriptions and no
standing policy, so agents knew the tools existed but not when to reach for
them. Lead each of the five descriptions with its trigger (when to use, when
not to) and add a task-tracking convention to .correx/project.toml, which is
rendered into every stage's L0 context.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Ranked exact/substring search: TaskSearch (all terms AND, ranked title >
key > goal > criteria > notes) over one project or the whole board via
TaskService.search. Surfaced as GET /tasks?q=, a read-only task_search tool
for agents, and a `/` filter in the TUI board.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
When an agent creates or claims a task via the tool, record its session as
a CONTEXT/SESSION link so the context bundle's session resolver has real
edges to inline (status/intent/workflow). Shared linkOriginSession helper;
reducer dedupes so create+claim from one session is a single edge. The REST
create path stays unlinked (no agent session there).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The context bundle resolved only TASK and DOC link targets; ARTIFACT and
SESSION stayed raw. Add two ports (TaskArtifactResolver/TaskSessionResolver)
with apps/server adapters: artifacts resolve to producing stage/session +
content excerpt from CAS, sessions to status/intent/workflow. Unresolved
targets still fall back to raw links. Wired through the tool and REST bundle.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
CRUD, lifecycle transitions, work-graph links and notes over /tasks;
share link-kind inference (TaskTargetKinds) between the REST route and
task_update tool. End-to-end route tests.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds task tracking as a first-class event-sourced aggregate (ADR-0012). The
event log stays the only source of truth; the board is a projection, and all of
a project's task events live in one stream (tasks:<projectId>).
core:
- new core:tasks module mirroring core:sessions: TaskStatus/State/Reducer/
BoardProjector/CounterProjection/Repository + TaskService write path that
appends events via the existing EventStore (no new storage)
- task events in core:events (sealed marker TaskEvent), registered in
Serialization.kt; status is derived by the reducer from lifecycle facts
- work-graph edges: TaskLinkType + typed TaskTargetKind on the link, so bundle
resolution dispatches on the recorded kind instead of guessing from the id
- delete is a soft tombstone (TaskDeletedEvent); cancel stays a visible outcome
surfaces:
- agent tools (Tier T2 mutators, T1 read): task_create / task_update /
task_delete / task_context, registered in both the main and per-workspace
tool registries
- REST GET /tasks/{id}/context for external agents
context bundle (TaskContextAssembler):
- task fields + acceptance criteria + relevant files + resolved dependency
tasks (live status) + raw related links + notes, with a compact render()
- semantic enrichment via a TaskKnowledgeRetriever port bridged to the existing
L3 retriever (late-bound holder for composition-root ordering)
- ADR/doc resolution via a TaskDocumentResolver port + filesystem adapter
(docs/decisions/adr-*.md, padding-agnostic; *.md paths)
Unit-verified across core:tasks / infrastructure:tools / apps:server; full
gradlew check green. Not yet live-QA'd end-to-end against a running server.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two reviewer-reliability tracks, both deterministic and unit-verified (no model/network).
§B-§5 — static_check stage seam. The StaticAnalysisRunner + StaticFindingsRecordedEvent +
reviewer-context filter already existed; what was missing was a stage that runs the tool and
emits the event. Added:
- StaticCheckStageExecutor: reads stage metadata (static_tool/static_argv), runs the configured
command via StaticAnalysisRunner, returns a StaticFindingsRecordedEvent. No-op (empty findings)
when no runner is wired or no command is configured — safe to carry unconfigured.
- A deterministic-stage seam in DefaultSessionOrchestrator.enterStage: any stage with
metadata["stage_type"] == "static_check" is run by the executor instead of the LLM subagent,
then advances on its (unconditional) exit edge.
- TomlWorkflowLoader: stage_type/static_tool/static_argv fields → StageConfig.metadata.
- role_pipeline.toml: a static_check stage between implementer and reviewer (no-op until
static_argv + a CommandRunner are set; activation is live-QA-gated, see StaticAnalysisRunner doc).
§B-§6 — critique-outcome producer. The CritiqueFinding type + CriticCalibrationProjection existed
but nothing fed them. Added:
- CritiqueFindingsRecordedEvent (+ CritiqueVerdict): the producing side — a critic's findings +
verdict for one review iteration, carrying modelHash for per-model calibration.
- CritiqueOutcomeCorrelator: pure loop-resolution logic deciding UPHELD (fixed between rounds) /
DISMISSED (persisted into an approved final) / INCONCLUSIVE (open at a non-approved terminal),
per critic (role + modelHash) and per finding id.
- A hook in completeWorkflow/failWorkflow that correlates recorded findings into
CritiqueOutcomeCorrelatedEvents at loop resolution — no-op when none recorded, idempotent.
(LLM-side finding emission stays a separate model-gated activation.)
Tests: StaticCheckStageExecutorTest (4), StaticCheckStageTest integration (1, fake runner →
event + transition), CritiqueOutcomeCorrelatorTest (8), CritiqueCalibrationWiringTest integration
(1, seeded findings → outcomes at completion), updated RolePipelineWorkflowTest. Full suites for
core:events/kernel/critique, infrastructure:workflow, testing:integration green; detekt clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
PlanStage gains a `writes` field (planner-declared workspace-relative paths/globs);
PlanDerivedManifest.deriveAllowedWrites + combine union it with the static baseline.
ExecutionPlanCompiler wires the union into StageConfig.writeManifest, enforced by the
existing ManifestContainmentRule via the same glob matcher (no parallel matcher).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- promote research fetch quality + content_sha256 from ToolExecutionCompletedEvent
metadata to first-class SourceFetchedEvent / LowQualityExtractionEvent (registered
+ serialization test); existing metadata write kept intact (additive)
- emitted from SandboxedToolExecutor on research-fetch tool completion, guarded on
the url/content_sha256/quality markers; reuses the extractor's LOW_QUALITY marker
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
First slice of the plan-generation pipeline epic. A pure-Kotlin PlanLinter runs over the
compiled freestyle ExecutionPlan (WorkflowGraph) before it is locked, complementing
ExecutionPlanCompiler (which throws on unreachable/unknown-kind/bad-edge) with the checks
it does NOT make:
- HARD unproduced_need — a stage needs an artifact no stage produces. Real gap: only the
TOML loader checked this; the freestyle compiler did not, so such plans compiled and
failed at runtime.
- HARD trap_state — a stage with no path to the terminal (inescapable loop / dead end).
Uses terminal-reachability, NOT "any cycle", so legitimate verdict-gated review loops pass.
- SOFT stage_count / fan_out / empty_brief / duplicate_brief — scored, never blocking.
FreestyleDriver lints after compile and records PlanLintCompletedEvent (pure function of
the recorded plan → replay-safe, no env observation in v1); a hard failure rejects the plan
(ExecutionPlanRejectedEvent source="lint") before the operator is asked or the plan locks,
so a broken plan never executes. Pays off now for single-plan freestyle; becomes the
per-candidate filter when multi-candidate generation (Slice 2) lands.
Deferred to later slices: file-path grounding + symbol resolution (needs an index),
token-budget/ADR-bypass checks, and a dedicated :core:planning module.
Run operator-configured static-analysis commands (compiler/detekt/formatters) as a
deterministic harness step on the producing stage, before the LLM reviewer. A stage
declares `static_analysis = [commands]`; after it produces its artifact, each command
runs in the workspace root and the results are recorded in a StaticAnalysisCompletedEvent
(invariant #9 env observation — recorded live, folded on replay, never re-run). A
non-clean command fails the stage retryably with its output fed back verbatim (§2:
compiler/test output is ground truth), so the implementer fixes it and only static-clean
code ever reaches the reviewer. This makes §5's "reviewer context excludes static findings"
mechanical — the findings are resolved upstream, so the reviewer can't waste inference on
what detekt catches for free; no output-parsing, no context plumbing.
- StaticAnalysisCompletedEvent + StaticAnalysisFinding (core:events) + registration.
- StaticAnalysisRunner seam + ProcessStaticAnalysisRunner (whitespace-split argv, stderr
merged, timeout→nonzero exit) in core:kernel; wired (nullable) into OrchestratorEngines
and the server. Null runner / no workspace root → logged no-op, never blocks the run.
- StageConfig.staticAnalysis + `static_analysis` TOML field; runStaticAnalysis folded into
a runPostStageGates sequence (produces → grounding → echo → static analysis).
- Documented `static_analysis` on the role_pipeline implementer stage (commented; commands
are workspace-specific). The reviewer prompt half shipped in 3467826.
New brief_echo stage before the planner in role_pipeline: the model restates
the analyst brief as a structured artifact, and a deterministic gate
(checkBriefEcho, opt-in via brief_echo=true) diffs the restatement against the
original brief. On divergence — dropped requirements (Jaccard coverage < 0.34)
or hallucinated files — it emits BriefEchoMismatchEvent and fails the stage
retryably, so a misread brief never reaches plan generation.
The diff (BriefEchoDiff) is a pure function of two recorded artifacts, so it is
replay-safe and records no observation; the event fires only on mismatch, for
audit. Symbols are recorded but non-blocking in v1. Mirrors the groundBrief-
References gate. role_pipeline only; freestyle_planning wiring is a follow-up.
Runtime install (like research): copy brief_echo.json + brief_echo.md and the
[[artifacts]] block into ~/.config/correx.
The rubber-duck path emits a fenced {"ideas":[…]} block; Ideas.parse strips it
from the chat turn and onUserInput records an IdeaCapturedEvent each (CHAT only).
IdeaReader folds eventStore.allEvents() into the active cross-session board, and
DefaultRouterContextBuilder injects a capped "## Ideas" entry — router-only, so
workflow stages never see it. Extended SYSTEM_PROMPT option (d) to capture ideas.
When given a directory path, file_read now returns its entries (sorted,
sub-directories suffixed '/') instead of failing. The tool description
and path-param schema advertise this so the model discovers it rather
than giving up after a missing-file probe. Generalises file_read into a
'read this path' primitive; the existing PathJail/Plane-2 containment
already covers directories.
Makes the research feature runnable end-to-end, off by default.
- config: [tools.research] (enabled, searxng_url, max_results, max_fetch_bytes).
- registration: web_search/web_fetch are built into BOTH the default and per-workspace
tool registries when research.enabled, sharing one HTTP client threaded from Main
(none built on the static path). Egress stays harness-enforced: web_fetch is T2
(operator-approved) and the existing NetworkHostRule still applies.
- workflow: examples/workflows/research.toml — decompose → gather → report, with the
three artifact schemas and prompts. Fan-out (search per sub-question, fetch per source)
runs as repeated tool calls inside the gather stage (Correx has no parallel agents);
per-source synthesis into the dossier is the compression step, so the report stage
consumes summaries, never raw pages. ResearchWorkflowTest validates the graph contract.
To run: set [tools.research].enabled, register the 3 [[artifacts]], copy research.toml +
prompts + schemas into the workflows dir, start SearXNG. Launch like any workflow (the
T2 fetch approval surfaces as an approval card; the report opens in the artifact viewer).
Follow-ups (noted, not blocking): batch fetch-approval at the source-list level (§3),
a dedicated SourceFetched/LowQualityExtraction event (quality + content hash are already
in tool-result metadata), dynamic per-session egress allowlist, and the web approval client (§6).
The research feature is "no new architecture" (spec §1) — composition of existing
mechanisms. The only genuinely new code is the two web tools + the HTML→markdown
extractor (spec marks them "(new)"), and those belong with ShellTool/FileReadTool
in :infrastructure:tools, not in a separate module.
Moves extract/ + web/ from the short-lived :infrastructure:research module into
:infrastructure:tools (packages com.correx.infrastructure.tools.{extract,web}); adds
jsoup + ktor-client there; removes the module from settings.gradle. No behaviour change
— 44 tests green.
Searches the self-hosted SearXNG JSON API: local-first, no API keys, no third-party
telemetry. Hits only the configured endpoint, so unlike web_fetch there is no
arbitrary-host egress to police — T1. Returns the top-N results as a readable list
(title / url / snippet) with the result URLs in metadata so the fetch stage can act on
them and the operator can approve the source list rather than each URL.
Tested with Ktor MockEngine (query encoding, result parsing/cap, empty + error paths).
Completes the §3 "two tools" pair; with slice-1 extraction this is the deterministic
+ network-tool foundation of the research workflow.
The first tool that punches through the sandbox to the network. T2, declares
NETWORK_ACCESS with url marked NETWORK_TARGET, so egress is enforced by the harness
(NetworkHostRule, plane 2) — not tool courtesy. Header-first guards (§3): content-type
checked before the body is read (binary/media rejected at the header via the extractor's
shared content-type policy); declared Content-Length capped; the read itself is bounded
so a server lying about length can't blow the cap (default 10 MB).
On success returns the extracted markdown plus metadata for citation lineage + replay:
url, content_sha256 (over the raw bytes), fetched_bytes, extractor_version, quality.
Consumes slice-1's HtmlMarkdownExtractor. Tested with Ktor MockEngine.
NOT yet registered in the live tool registry: with an empty networkAllowedHosts the
host rule allows everything, so exposing a fetch tool before the egress allowlist
(SearXNG endpoint + approved-source hosts, per-session) is wired would be an open-egress
hole. Registration + egress config land with the research workflow slice.
First slice of the research workflow: the pure, zero-inference extraction pipeline
that the WebFetchTool will consume. New :infrastructure:research module (jsoup).
HtmlMarkdownExtractor: content-type dispatch (HTML extracted; JSON/text pass
through; binary/media/pdf rejected at the header), then structural strip
(MainContentSelector: drop script/style/nav/footer/aside/form + class/id blocklist)
+ density extraction (text − link-text score, semantic <article>/<main> preferred)
+ DOM→markdown (MarkdownRenderer/Table/Inline: headings, lists incl. nested, tables,
fenced code, blockquotes, links, emphasis). Output below the min-length threshold is
flagged LOW_QUALITY so the workflow can route around dead sources (SPAs, paywalls).
extractorVersion ("html-md-1") is pinned on every result so replay reads recorded
artifacts and never re-extracts — the embedding-hash discipline (§5, ADR-0000 §10).
Next slices: WebFetchTool + egress policy + CAS wiring (emits the fetch +
LowQualityExtraction events), then WebSearchTool/SearXNG, workflow graph, synthesis.
Local analysts hallucinate file paths the same as planners. A stage can now
opt in with `ground_references = true`: after it produces its brief, every
workspace-relative file path the brief named is checked for existence, and a
path that doesn't exist fails the stage (retryable) with the misses fed back
("Reference only files you have read") so the model re-grounds.
Grounding probes the filesystem directly rather than the repo map — the map
is recency-ranked and depth/extension-limited, not a complete existence
oracle, so "absent from the map" can't be a hard failure. Existence is a
nondeterministic observation, so the result is recorded as a
BriefGroundingCheckedEvent (invariant #9); replay reads it back and never
re-probes.
- BriefGroundingCheckedEvent + GroundingReference, registered in eventModule
- BriefReferenceExtractor: pure, deterministic pull of relative file-path
references from a brief artifact JSON (needs a '/' and a dotted extension;
skips modules, bare names, URLs, absolute/parent paths — so coarse
"subsystem" mentions the brief is allowed to make aren't false-flagged)
- SessionOrchestrator.groundBriefReferences on the post-verifyProduces path,
gated by stage metadata; uses the existing worldProbe
- TomlWorkflowLoader `ground_references` → metadata; role_pipeline analyst
opts in
Scoped to file references with a '/' + extension (not bare basenames, which
a depth-limited probe can't resolve), so a flagged miss is high-confidence.
The symbol-grounding half of §3 is left for a real symbol index.
Scope creep — an implementer writing files it never declared — is the
dominant local-model failure that compiler/tests don't catch. A stage can
now declare a `writes` manifest (workspace-relative globs); a FILE_WRITE
that resolves inside the workspace but outside the declared set is raised
as PATH_OUTSIDE_MANIFEST → BLOCK, so the model gets the violation as tool
feedback and retries within scope (the §2 bounded-retry signal).
Implemented as a plane-2 rule beside PathContainmentRule (same effect-based
dispatch on FILE_WRITE, same symlink-safe WorldProbe resolution, same
recorded observations so replay reads them back — invariant #9). An empty
manifest is unrestricted; out-of-workspace targets stay PathContainmentRule's
concern, so the two rules don't double-flag.
- StageConfig.writeManifest + TomlWorkflowLoader `writes` parsing
- ToolCallAssessmentInput.writeManifest, threaded from the active stage via
SessionOrchestrator.runPlane2Assessment
- new ManifestContainmentRule, registered in the server assessor
- shared candidatePathStrings extracted so both path rules judge the same
target set
- role_pipeline.toml documents the option on the implementer stage
The dedicated ManifestViolationEvent the spec names is not added: the
violation is already recorded replayably in ToolCallAssessedEvent (issue +
observations + BLOCK) and surfaced in the TUI rationale band. A first-class
event is worth adding only once reconciliation consumes it.
Live repro: architect emitted a verify stage with kind=analysis and a
verdict-equals edge. The kind schema doubles as the LLM response format,
so 'verdict' was never emitted and the workflow failed at its final hop
at runtime. The compiler now requires artifact_field_equals fields to be
declared by the producing stage's kind schema, failing the plan at lock
time with a pointer at review_report. Architect prompt rule added;
reachability check extracted alongside the new validation.
Live repro: the architect wired all three stages straight to 'done'
instead of chaining them, so phase 2 legitimately completed after
stage 1 of 3 — a silently truncated plan reported as success. The
compiler now walks the edge graph from start and fails compilation
when any declared stage is unreachable, surfacing the broken topology
as ExecutionPlanRejected at lock time instead. Architect prompt edge
rules rewritten to demand explicit chaining (the old wording was
garbled).
ExecutionPlanCompiler treated the plan's single 'produces' string as a
registry kind id, while the architect prompt described it as a unique
artifact id referenced by needs/edges — the model followed the prompt,
invented descriptive ids (files_created), and every freestyle plan
failed to compile. Collapsing both onto one string is also unsound:
two stages producing the same kind would collide on slot name and make
artifact_validated edge conditions ambiguous.
Plan stages now carry an optional 'kind' selecting the registered kind
(mirroring TOML's produces = { name, kind }); 'produces' stays the
unique slot name. Missing 'kind' falls back to the old produces-as-kind
reading so existing plans still compile. Vocabulary entry, architect
prompt, and execution_plan schemas updated to match.
LlamaCppTokenizer posted {"tokens":[text]} but llama.cpp's /tokenize expects
{"content":text}; the server silently answered {"tokens":[]}, so countTokens
returned 0 for every string. Every context entry carried tokenEstimate=0 and
the whole budget/trim pipeline was blind — live QA saw a 34k-token prompt
sail through a 16384 stage budget (three raw file_read results of plan docs),
crater t/s, degrade output, and fail the stage on validation 3x.
Also guard estimateTokens: a zero count for non-blank content falls back to
the chars/4 heuristic so a lying tokenizer can never blind budgeting again.
- Add existsByTurnIdPrefix(prefix) to L3MemoryStore interface; implemented in
InMemoryL3MemoryStore (mutex-guarded) and TurboVecL3MemoryStore (metadata map).
All test fakes (TrackingL3MemoryStore, CapturingStore) updated.
- Introduce RepoMapIndexerPort interface; RepoMapIndexer implements it. Allows
injection of test doubles without framework overhead.
- ProjectMemoryService.indexAndRecord: reads latest WorkspaceStateObservedEvent
from event store (invariant #9 — recorded fact, not re-probe); skips scan+embed
when L3 already holds entries for "repomap:<root>:<stateKey>"; passes stateKey
to RepoMapComputedEvent; embeds each scanned entry after append; per-entry
failures warn-logged (CancellationException rethrown).
- Tests: InMemoryL3MemoryStoreTest +2 cases; new ProjectMemoryServiceReuseTest
covers same-key skip, changed-key re-index, no-observation always-reindex,
and embedding turnId tag verification.
- generic SystemResourceProbe: owner-agnostic /proc/meminfo system RAM
overlaid on the vendor GPU probe, wired on both managed and static paths
so the VRAM/GPU/RAM gauge renders with an external llama-server (was dormant)
- F-002: classify provider 4xx (e.g. llama.cpp 400) as terminal, not retryable
(except 408/429); stops retrying deterministic request failures to exhaustion
- F-003: FileSystemPromptLoader expands leading ~ / ~/ to user home
- F-004: map InferenceFailedEvent + RetryAttemptedEvent to new
ServerMessage.InferenceFailed / RetryAttempted, decoded+rendered in tui-go;
ArtifactContentStoredEvent explicitly mapped to null (internal, no warn)
- tests: tilde expansion, SystemResourceProbe (static/managed/fail-soft)
Live QA audit of role_pipeline against a sample repo surfaced and fixed a
chain of defects that prevented any tool+artifact workflow from running
end-to-end. The pipeline now completes cleanly with a real, validated,
reviewed file change.
- F-008 workflow: propagate stage token_budget -> generationConfig.maxTokens
(TomlWorkflowLoader) so large stages stop truncating at the 2048 default.
- F-009 tools/kernel: file_read missing-file is recoverable; recoverable tool
errors feed back into the loop instead of aborting the stage.
- F-010/F-018 kernel: reject premature stage_complete and nudge the model to
emit the owed artifact / invoke a write tool (bounded by MAX_TOOL_ROUNDS).
- F-011 schemas: list-typed artifact fields string -> array (analysis/design/
impl_plan) to match model output.
- F-012 kernel: strip an outer markdown code fence from LLM artifacts.
- F-013 validation: payload-validation failures are retryable; structural
failures stay terminal (wires the invariant-#7 reject+retry contract).
- F-014 tools: file_edit schema matches its real params; edits emit a diff.
- F-015 kernel: record tool-execution events + materialise a real file_written
artifact from the on-disk write; gate validation so a no-write stage cannot
be rubber-stamped (no more false-success runs).
- F-016 server: raise stageTimeoutMs 60s -> 180s for slow local models.
- F-019 artifacts/kernel: file_written artifact carries the unified diff so the
reviewer can verify the change (shared DiffUtil; file_write emits a diff too).
- F-020 tools: model-correctable file_edit failures are recoverable + steer
toward replace/file_write over the fragile patch op.
Full findings register (F-001..F-021, incl. open F-021 resume rehydrate bug)
in qa/audit-report-2026-06-08.md. All module tests green.
Two fixes surfaced by the 2026-06-08 QA audit running role_pipeline on a local
llama-server.
F-001: LlamaCppInferenceProvider sent both a GBNF grammar and tools on the same
request; llama.cpp rejects the combination ("Cannot use custom grammar
constraints with tools"), failing every stage that has tools and produces a
schema-constrained artifact. Drop the grammar when tools are present and rely on
post-hoc schema validation + retry (invariant #7 still holds). TODO left for a
first-class emit_artifact tool as the proper fix.
F-005: FileReadTool resolved relative paths via toAbsolutePath() against the JVM
process cwd instead of the bound workspace, so the file-read jail anchored to the
wrong root and disagreed with the Plane-2 PATH_CONTAINMENT check. Add workingDir
to FileReadConfig, wire workspace.workingDir in the per-workspace tool builder,
and give FileReadTool a resolvePath() mirroring FileWriteTool/FileEditTool.
The resource gauge was NVIDIA-only — on an AMD/ROCm box Main fell back to
UnavailableProbe, so the TUI showed no VRAM. Add AmdResourceProbe backed
by `rocm-smi --showmeminfo vram --showuse --csv`: it skips the warning
preamble, locates the header by column name (order/version tolerant), and
converts the byte VRAM figures to MiB. Process RSS reuses /proc/<pid>/status
(only populated for a managed-model pid). Main now selects NVIDIA → AMD →
Unavailable. Tests cover the real ROCm 4.0 CSV + reordered columns.
kami