Resolve the run workspace's environment with workenv (direnv, parent
.envrc, mise; same account rule as Claude chats) at run start and resume,
and apply it to Pi stage processes, a per-environment Pipelines-owned
OpenCode stage service (keyed by a hash of the change, reused across runs,
idle ones stopped after PIPELINES_OPENCODE_STAGE_IDLE_TIMEOUT, revived per
registration), delivery-gate checks and merges, plain-stage opencode CLI
calls and file-change capture git reads. Record the directory, .envrc
files, sources and variable names (never values) on the run detail; a
blocked .envrc becomes a run warning.
The prompt now says output.md must start with # outcome and the outcome
word alone on the next line, then # handoff, then detail. Runs 241/243
followed the old "start with # handoff" wording and dropped the outcome.
When a completed turn lacks a valid outcome, the OpenCode/agent and Pi
clients send one short follow-up in the same session asking only for the
outcome line. An outcome-only reply goes at the top of the original report;
a still-invalid reply fails with the same validation error as before. The
follow-up's tokens accrue to the session usage. No outcome is inferred.
Missing tools/evidence now name blocked when declared, else error. The
parser tolerates a trailing period on the outcome word.
A stopping server no longer interrupts harness sessions. Runs are canceled
with a shutdown cause, so stage clients leave admitted sessions running and
owned; start-up recovery re-adopts live OpenCode V2 stage sessions (wait,
read back the result, continue the graph), restarts unadmitted visits,
continues runs between stages, and falls back to interrupted with a reason
when the harness is gone. Chats with live work are recorded on shutdown and
reattached on start; turns that ended during the downtime are folded in
from history. The server-spawned V2 chat serve is now a detached owned
service. PIPELINES_SHUTDOWN_SESSIONS=stop restores the old behaviour.
Stages now run on a detached Pipelines-owned opencode serve started with
OPENCODE_CONFIG_CONTENT carrying stage tool_output (16384 bytes / 400 lines)
and compaction.auto defaults, configurable through PIPELINES_OPENCODE_STAGE_*.
Chats keep their existing service and configuration; a server-spawned chat
serve can take PIPELINES_OPENCODE_CHAT_CONFIG. The service shares the user's
OpenCode config, credentials and session database, keeps a stable private
registration so retained references (recovery, interrupt, export) stay valid,
revives on the same port after a crash, replaces itself when idle after a
settings change, and is stopped on shutdown only when idle.
PIPELINES_OPENCODE_STAGE_SERVICE=shared restores the previous behaviour.
Every stage used to be told to read memory.md, the full output of every
earlier turn (19 KB average, 92 KB max), and paid for it on each model call.
memory.md is now essential memory, hard-capped at 4 KB and inlined in the
prompt: the task, the previous stage's # handoff section (or its truncated
output head), a one-line-per-stage outcome ledger and the server-captured
changed files. history.md keeps the old full memory on disk; read-capable
stages are pointed at it and input.md for on-demand detail, and it is never
inlined. The prompt asks each stage for a short # handoff section.
The server skip fallback for stages with a skipped edge now applies only
when every earlier visit recorded a complete file-change capture with no
changed files and no image artifact. Any change (no repo-specific UI path
or extension list), any partial/unavailable/missing capture, a human
decision or a checkpoint-retry prefix runs the model. auto_skip=false opts
a stage out. The fallback verdict names the server fallback.
Capture compares over-budget text by hash and, in Git, lists paths it
could not compare when their git status changes, marked no_patch, so a
large workspace no longer records files: [] for a real edit. Run detail
gains a warnings list for partial/unavailable captures, also printed by
inspect on stderr.
Persist three global ntfy switches in server_settings: run_finished
(default on: auto-mode runs notify on success, read inside the
transition transaction; on/off modes unchanged), chat_needs_you
(default on: gates chat question/approval alerts) and
chat_reply_finished (default off: one alert per finished assistant
turn). Expose notification_settings / set_notification_settings over
the websocket protocol, add a Phone notifications section with
Material and Cupertino checkboxes in Settings, document the settings,
and keep the cli, tui, stage and pipeline test packages from
inheriting a real ntfy configuration.
Run notifications now always click through to pipelines://run/<id>; a valid
PIPELINES_WEB_PUBLIC_URL adds an "Open in browser" ntfy view action instead
of replacing the click. Queued rows derive links from the run id at delivery,
so legacy rows holding a web URL still open the app.
Unresolved chat questions/forms and permission requests publish one async,
bounded notification to the system destination (pipelines://chat/<id> plus the
browser action), deduplicated per form/request id across reconnects and
replays. stage.PublishNtfy gains optional, strictly validated view actions.
Server tests now clear the operator's real ntfy environment.
Six computed Flutter runs stopped at the export handoff gate because the
rendered session.html cap was 2 MiB while their complete sessions (well inside
the 8 MiB source read bound) rendered between 2.25 and 4.28 MiB once html
escaping and the readable projection duplicated native text.
Raise SessionExportFileLimit to 8 MiB, the server's existing per-artifact file
bound, so a complete export is still listed and served. Normal export
verification, versioned repair and the handoff gate keep using the single
shared constant. The source read bound is unchanged, and an over-cap render is
still visibly truncated and recorded partial.
Add focused regressions for a complete export above the old cap, an over-new-cap
file rejected as partial, repair above the old cap leaving original evidence
untouched, the gate accepting the larger repair, and the constant staying inside
the artifact file bound. Update the artifact, handoff-gate and operations docs
and record the diagnosis in docs/notes/session-export-render-cap.md.
A probe stage read "public/" in its task, created a workspace-relative public/
folder, and also copied to the run's shared path. The prompt now states that the
absolute path is the shared folder and a relative public/ is a different,
unshared directory.
The successful retry in TestDefaultRetriesAndTimeoutResetEachAttempt was
racing its own 20ms deadline when the full suite ran in parallel. Keep the
two-attempt behavior and give the retry realistic headroom.
Each model visit now gets <run>/stages/<seq>-<stage>/ with input.md,
memory.md, output.md, private/ and a shared run-level public/. The model
reads the two input files, follows the stage's instructions verbatim and
returns complete markdown; only the # outcome section is parsed.
- add internal/turn layout, render and output parsing
- write input.md/memory.md before a visit and output.md from the final
message; inventory all turn files as stage artifacts
- accumulate completed turns and system-written public file additions in
memory, including across retries and resume
- remove the JSON envelope, strict field validation, line format and the
repair ladder for model stages
- republish pipeline authoring prompts and docs in the new style
Checks: go build ./..., go vet ./..., go test -count=1 ./..., gofmt -l .
Drop the resource allocation feature end to end: store tables and migration, runner scheduling and cleanup, protocol types and commands, CLI commands, TUI screens, tests and docs. Existing databases open cleanly and definitions carrying a retired resources block still validate.
Add validated OpenCode-only delivery capabilities, canonical linked-worktree access checks, isolated candidate verification, and guarded main settlement.
Wire the bounded reconnect-foreground controlled HTTP fixture through ordinary server construction so target verification obtains server-controlled provenance and rejects unknown or incompatible fixture evidence.
Checked packaged templates and packaged authoring skill references; no updates are needed because this draft-only infrastructure activates no definition. Recorded the stale reference and activation guidance in docs/notes/delivery-stabilization.md for the docs fold.
Add durable FIFO capacity-one resource leases for executable stages, including cleanup-confirmed release, conservative restart quarantine, queue state events, and CLI/TUI/Android waiting presentation. Add scheduler, cleanup, protocol, client, and Android coverage plus a proposed Android v3 resource draft.
Checked packaged templates and authoring skill references: neither needs updating in this item. The per-item docs note records the shared documentation that docs-fold must update; no shared README or packaged reference was edited.
Derive validated same-run visit artifact capabilities in the runner and pass them to OpenCode and Pi clients. Permit supported file tools to read completed prior visits and write only the current visit, preserving workspace, secret, database, and shell restrictions.\n\nChecked packaged templates and authoring skill references: neither needs updating in this item; docs/notes/scoped-artifact-access.md records the shared documentation staleness for the docs-fold writer.
Identify configured attempt expiry during V2 active-session idle waits and retain bounded phase, timeout and session diagnostics without claiming a lost response. Preserve genuine operation failures, error identities, redaction, cleanup and admitted-work replay protections. Add stage regressions and scratch-server CLI/websocket retention coverage.
Checked packaged templates in examples and the pipeline template generator: none needs updating because timeout configuration and contracts are unchanged. Checked packaged authoring skill references: debugging.md needs a STAGE DEADLINE playbook and clarified transport guidance. Recorded that stale shared documentation in docs/notes/stage-timeout-diagnostics.md for docs-fold; shared documentation was deliberately not edited.
Verification: go build ./...; go vet ./...; go test ./...; go test -count=1 ./internal/server ./internal/stage ./internal/pipeline ./internal/protocol; go test -count=1 -run TestStageDeadlineDiagnosticThroughCLIAndWebsocket -v ./internal/server; gofmt -l .; git diff --check. All passed. Live-provider behavior not tested; no deploy or deployed-service restart.
A stage's session was exported twice, as session.md and session.html, and both
files were presented. It is now one html file; the markdown export is retired.
- internal/stage/session_export.go: ExportSession renders and writes exactly one
file, session.html, via writeSessionFile; renderSessionMarkdown and its helpers
are gone; SessionExportResult carries HTML/HTMLBytes only. The source read, the
8 MiB source / 2 MiB rendered bounds, the truncation banner, the
never-overwrite rule and the honest empty-session handling are unchanged.
- internal/protocol/runs.go: SessionExport drops markdown/markdown_bytes (a
record written before the change still decodes; unknown fields are ignored),
and adds LegacySessionMarkdownArtifact with RunStep.PresentedArtifacts(), the
one Go rule that hides the retired session.md from every artifact listing.
- internal/server/runner.go: evidence.Markdown/MarkdownBytes are no longer copied.
- internal/tui: the artifacts panel, the recorded count and the per-visit summary
derive from PresentedArtifacts(), so the terminal never lists session.md.
- web/src/app.js: the spike's renderArtifacts applies the same filter.
- android: the Session section is the html export alone in a WebView — SessionView,
the Markdown/HTML toggle and the session markdown path are removed;
presentedArtifacts(step) filters the retired session.md out of the per-visit
gallery, the run gallery and its count, the verdict/selection lists and the
timeline indicator; data/protocol SessionExport drops the markdown fields.
- Tests pin the single html file, prove no session.md is written and that a
pre-existing one is neither overwritten nor presented, and prove a legacy
export record still loads and its session.md is hidden.
The legacy session.md of an older run stays on disk (dated evidence); it is only
no longer produced and no longer presented.
Checks (macOS, Go 1.24, Android Gradle 8.11.1, 18 Oct 2026): gofmt -l . prints
nothing; go build ./... && go vet ./... clean; go test ./internal/... ./cmd/...
-> 19 packages ok, 0 failures; node --check web/src/app.js ok and node --test
web/test -> 20/20; gradle --offline --no-daemon :app:assembleDebug
:app:testDebugUnitTest -> BUILD SUCCESSFUL.
Packaged references checked: internal/author/skill/references/delivery.md and
operations.md both described the session.md+session.html pair, so both are
updated here to describe the single session.html export and the retired, hidden
session.md. README.md and internal/author/skill/SKILL.md carry no session.md
reference and install.go's managed inventory is unchanged (no reference added or
removed). docs/notes/session-html-only.md records the remaining stale shared
documentation (docs/stage-artifacts.md, docs/app-integration-verification.md) for
the docs-fold writer; shared docs are not edited by this item.
Pure move: split the three multi-feature server files into per-concern files,
with no behaviour change - same package, same exported and unexported
identifiers, same call graph, same tests.
- internal/server/chat.go -> chat.go (broker seam, client token, dispatcher),
chat_hub.go, chat_claims.go, chat_permissions.go, chat_sessions.go,
chat_stream.go, chat_commands.go
- internal/stage/model.go -> model.go, model_run.go, model_failure.go,
model_validation.go
- internal/pipeline/runs.go -> runs.go (read side), runs_schema.go,
runs_create.go, runs_mutate.go, runs_sessions.go
Checked before committing:
- Packaged templates and internal/author/skill/references: neither refers to
internal/server/chat.go, internal/stage/model.go or internal/pipeline/runs.go,
so nothing there needed updating (the change is a move with no behaviour
change).
- gofmt -l on the three packages: clean. go build ./..., go vet ./... and
go test ./... pass. go.mod/go.sum unchanged.
- No test, docs, android/, install or Docker files touched.
Known stale path pointers deliberately left to the single docs writer (shared
docs) and out of scope (android/): docs/chat-verification.md and
android/README.md name the old file paths.
Carry the usage a model stage already measures for its budget
(stage.Budget.Snapshot()) onto the step as typed evidence and total it in
one pass, rather than a second harness read, a transcript parse or an
output-length estimate. A shared budget across stage revisits records only
the visit's increment, and a revisit that observes no new report records no
usage instead of a fabricated known zero.
- protocol: StageUsage/RunUsage/TotalUsage and UsageTotals.
- pipeline: step_usage table + migration, COALESCEd into the one step
projection, FinishStep persistence, one json_extract history aggregate,
retention removes the rows with their run.
- server: stageUsage maps the budget snapshot, stamping per-stage model and
usage for every model stage (including failures).
- cli: inspect/stats human and JSON output.
- budget/pi: retain the reported in/out split, mark a total-only (Pi) report
split-unknown, and keep an unreported cost unknown (never zero).
Evidence and docs updated in the packaged skill references and docs/cli.md;
the live database has no step_usage rows yet (all history predates this), so
per-run figures are reported as unavailable, not zero.
Every OpenCode-harness model stage now writes its real session into its own
step artifacts directory as session.md and session.html, so the existing
bounded artifact listing, the /artifacts/<run>/<step>/<name> serving path and
the artifact scan find and serve them with no special casing, while the stage
keeps its bounded preview in the run record unchanged.
- internal/opencode/export.go adds the one shared native export invocation and
the exact session-ID/messages validation used by both Connection.Messages and
the new ExportSession, so the readback path does not diverge. The backend's
published subcommand is used (v1 `export <id>`, v2 `session export <id>`),
the source read is capped at 8 MiB, and a larger session is read as a prefix
and reported truncated instead of failing. boundedOutput.ReadFrom keeps the
cap effective so overflow and the truncation path are real.
- internal/stage/session_export.go renders markdown and html with a metadata
header, every role and every part (unknown part types as raw JSON, every
session text escaped in html), atomic temp+rename writes that never overwrite
an existing file, a visible truncation banner at both ends when the source or
the 2 MiB rendered cap cuts it, and an honest 0-message header for an empty
session rather than fabricated text.
- protocol.SessionExport and RunStep.SessionExport are stored in a new
step_session_exports table (CREATE TABLE IF NOT EXISTS, so an old database
gains an empty table with no migration) loaded through the existing main
SELECT alongside the other per-step evidence and deleted by run retention.
- internal/server/runner.go exports after the post-stage filediff capture and
before scanArtifacts, with context.WithTimeout(context.WithoutCancel(ctx),
10s); it never waits for session idleness and a failure or a Pi stage (no
native export path) records a sanitized diagnostic and continues, never
changing the stage status, outcome, target, timing or result validation.
Reflected where it belongs: the packaged skill references
(internal/author/skill/references/operations.md and delivery.md) carry the new
Session export rules and the run-artifacts bullet, and docs/stage-artifacts.md
documents the names, the <run>/<step> paths, the served URL, the caps, the
truncation labels, failure isolation and the Pi limitation. The packaged
templates (examples/*.md) only tell a stage to save its own artifacts and say
nothing about the server-written session export, so none needed updating; no
new reference file was added, so install.go's managed inventory is unchanged.
Shared documentation: README.md's per-step artifacts sentence is now stale - it
should name the exported session alongside images, screenshots, reports and
logs. It is deliberately not committed here; the shared documents are folded by
the single documentation writer.
Checks: gofmt -l on the touched packages produced no output; go vet ./...
passed; go test -race -count=1 ./... passed.
Nothing told the operator that a run was waiting for a person or had failed; they
found out by going looking. A run that needs attention now reaches the ntfy they
already run, and the app answers "what needs me now" in one live, ordered surface.
Server (smallest change, reusing the existing publisher):
- internal/server/notify.go adds notifyRun, called from finishRunState (failed)
and when a run parks on a human (runner.go runLoop). It publishes to the same
already-configured `system` destination through the unchanged stage.PublishNtfy,
so there is no second publisher, topic or service and the token handling is the
stage's own. runAttentionEvent maps only waiting and failed; queued, running,
succeeded, canceled and interrupted publish nothing.
- No storms: runner.attentionSent records sequence:event per run before publishing,
so a run that stays waiting is never republished on a re-request or restart, and
a stage notify policy that already covers the same event marks it first so the
run-level default is suppressed. The set is dropped when the run is terminal. A
delivery failure is logged, never fails the run, and is not retried.
- The payload names the run and pipeline and carries a click action
pipelines://run/<id> (validated: absolute URL, no userinfo, no fragment, no
credentials). PublishNtfy gained the optional click parameter; existing callers
pass "".
App:
- AndroidManifest/MainActivity declare the pipelines://run/<id> deep link
(singleTop); runIdFromDeepLink parses it and PipelinesApp opens that run once,
clearing the link. It opens a run; it starts nothing.
- ui/Attention.kt is the pure model (status -> Decision/Failed/Finished or
nothing, ordered by kind then newest-first) and ui/components/AttentionSurface.kt
is the floating indicator plus bottom sheet. It is derived from the same run
events the server already pushes, so an entry appears with no refresh and clears
when resolved; a finished run clears once looked at, and any entry can be
dismissed, persisted in SettingsRepository. The existing destinations, pill and
bottom strip are untouched and no competing navigation surface is added.
Checked before committing:
- Packaged authoring skill references: internal/author/skill/references/
notifications.md documents the automatic run-attention notifications and the
app surface. The only packaged "template" carrying notify is that reference; no
pipeline definition template exists to update and install.go's managed file list
is unchanged (notifications.md already existed).
- Repository guidelines: docs/ntfy.md gains an "Automatic run-attention
notifications" section and no longer claims "No block means no notifications";
the top-level README phone-notifications note now says a waiting/failed run
notifies automatically; android/README.md gains the dated feature section and
supersedes the previous section's note that the attention surface was absent.
- docker.md, install docs and the CLI/TUI docs describe mechanisms this change
does not alter; left as they are.
Tests: go test ./internal/server/... ./internal/stage/... ./internal/author/...
all ok; notify_test.go covers the event mapping and the real HTTP delivery
(destination, topic, click, no duplicate). Android :app:testDebugUnitTest ->
BUILD SUCCESSFUL, 58 suites / 531 tests / 0 failures (new AttentionTest, 13 cases,
covering ordering, appearing from a push, clearing on resolution, dismissal, the
empty case, persistence and the deep link). Device screenshots and the phone's own
notification are visual review's per the brief's device policy.
Scope: android/ plus the smallest server change; no new dependency, no run started
from the app, no installer or Docker change.
Treat a failure to reach the harness before any prompt or tool call is
admitted (connecting or creating the session) as a distinct retryable
reach failure, so a transport blip at the first stage no longer ends a
mutating run; keep post-admission readback/poll/agent failures retried
only on an explicit read-only stage and never replay admitted tool work.
Preserve the true cause when the context is already done (join ctx.Err
with the typed transport error) and report a sanitized unknown provider
error name/status instead of a blanket summary.
Make the isolated format-repair deadline proportional to the stage
timeout (floored, capped and clamped to remaining minus a margin) so a
slow-but-repairable result is recovered without ever extending the stage
deadline; the bounded ladder, value guard, deny-all formatter session,
cumulative usage accounting and no-tool-replay rule are unchanged.
Update docs/model-results.md, docs/pipeline-performance-review.md and the
packaged skill references; add stage, opencode and server fixtures.
A planner had no machine-readable view of the tool set later stages would
have, so it could emit acceptance checks that required tools (shell) the
implementer was not permitted, causing a blocked run.
modelPrompt now emits exactly one authoritative line naming the stage's
permitted tools, derived from the definition via the new effectiveTools
helper. effectiveTools reuses ToolNames: a present tools block, including
an explicit empty one, is returned as parsed; an absent block yields the
legacy mutating read/glob/grep/edit/shell bundle so it is never
misrepresented as empty or read-only. The bundle is declared once as
legacyTools and reused by tools.go, agent.go, pi.go and permissions.go.
The redundant client-side strings.Replace of the tools line is removed
from AgentClient.Generate and PiClient.Generate; the agent-only guidance
sentence is preserved by appending it in agent.go, and Pi appends a short
mapping line derived from piTools. Edit-scope and permitted-shell text,
routing, review, budgets and retries are unchanged.
Adds internal/stage/model_prompt_test.go covering explicit, absent and
explicit-empty blocks plus derivation edge cases, and a guard that the
agent prompt still carries the tools, edit-scope and shell-command text.
Introduce a short, markdown-ish result format whose first non-empty line is
the declared outcome and whose "# output" section carries the routing data,
so stage output stops carrying multi-kilobyte prose inside a JSON envelope.
- New internal/stage/line_result.go: resultFormat() gates legacy JSON ('{'
first) vs the new line format; decodeLineResult() parses outcome-from-first
line, requires a JSON-object "# output" section, stores extra sections
verbatim under normalised keys, and records a reserved result_parse status
(parsed/partial/failed). Failures return a salvage Result plus
ErrInvalidModelResult, and never route an undeclared outcome.
- internal/stage/model.go: write the raw response to raw-output.md (0600,
best-effort) before parsing, and route the line format through
decodeLineResult while leaving the JSON path (stageResultText ->
decodeModelResult -> repairModelResult) unchanged.
- internal/stage/agent.go: agentResultText returns line-format text verbatim
so the "# output" body is not stripped by the JSON-object extractor.
- internal/server/runner.go: merge a non-empty salvage Result into the error
envelope under "salvage" on failure (a no-op for the legacy JSON path).
- internal/stage/line_result_test.go: raw-text fixtures covering happy path,
missing/undeclared outcome, missing "# output", best-effort extras,
adversarial prose, legacy JSON unchanged, >=64KB responses, and
raw-output.md written before parsing.
Refinements on top of the human decision stage (a6aff47):
- `pipelines decide` now reports the feedback the server actually stored for
the chosen option, so a repeated (idempotent) decision shows the recorded
evidence instead of the ignored request text. A repeated decision is covered
by a new assertion that also checks the step output does not change.
- The decision option/feedback output keys are defined once in protocol
(DecisionOptionKey/DecisionFeedbackKey) with a DecisionEvidence reader, and
the store and CLI use them instead of bare string literals.
- piEnvironment drops any inherited PIPELINES_ARTIFACTS_DIR before spawning Pi,
so the artifactsDir argument is authoritative and an ambient directory (for
example Pipelines' own when a stage runs inside another stage) never leaks
into a child that was given none. Covered by a test with an ambient value set.
A new registered stage type `human` (internal/stage/human.go) has no model,
agent, tools or configs: its required `prompt` block carries the operator
instructions and its non-error edges are the selectable options. It parses,
validates, renders, exports and imports like every other type (no parser
special-case, no config keys).
At runtime the runner parks the run instead of executing: the step and the run
persist `waiting` with self-contained {prompt, options} evidence, and the run
returns without starting later stages, capturing file diffs, scanning artifacts,
holding a goroutine/timer or applying a stage timeout. The wait is durable: a
restart leaves the waiting rows untouched (RecoverRuns excludes them) and a
rebuilt runtime over the same database still accepts the decision.
`pipelines decide <run-id> [stage] --option KEY [--feedback TEXT] [--json]` and
the TUI `c` picker resolve the decision from the pinned graph in one store
transaction, record the choice and feedback as the step's evidence and resume
along edges[KEY]. An unknown option is refused with the valid list, a stage that
is not waiting is refused, and a repeated decision for an already-decided visit
does not double-apply. Cancel and retry keep working while waiting and after a
decision; a human stage is never killed by a stage timeout or max_visits, and a
`notify` policy can fire an `on=waiting` event.
Docs: the authoring skill references (format.md, operations.md), docs/cli.md and
README describe the stage, the command and the TUI picker. Tests cover stage
validation, the parsing round-trip, the store transitions, a websocket wait/decide
integration, cancel/retry, waiting notifications and a persistence-layer rebuild
(second Open + newRunner over one database, no process restart).
A model stage whose final result was malformed failed the whole run even when
the work and evidence had survived, because the only recovery was an isolated
formatter call on a fixed 30-second deadline that could time out on its own
(run 55, ui_reviewer of pipeline 9 v6).
Add an ordered, deterministic local repair pass (internal/stage/result_repair.go)
that runs before any model call: (a) append missing closing delimiters via the
existing closeJSONContainers; (b) remove one unambiguous trailing comma before a
closing } or ]; (c) insert exactly the single comma/colon the JSON parser reported
at its byte offset. Every candidate must pass losslessSyntaxRepair(original,
candidate) and decodeModelResult before it is accepted, so no field value is
invented, removed, reordered or rewritten. Ambiguous payloads (an unescaped quote
or colon inside a string value, or a stray delimiter with no single completion)
return not-ok and fall through to the bounded model re-emit rather than guessing.
stageResultText still strips surrounding prose and a single code fence first;
model_prompt.go now also strips prose around a single balanced-but-malformed
object span so the deterministic pass sees just the object.
Replace the fixed 30-second formatter deadline with repairDeadline(payloadBytes,
remaining), which grows monotonically with payload size, has a floor, and is
clamped to the attempt's remaining budget minus a safety margin, so the stage
context remains the outer bound and a large result is not guaranteed to time out.
The one-request-per-attempt, 32 KiB bound and the deny-all, retained, new-session
property of RepairFormat are unchanged for OpenCodeClient and PiClient.
Record an explicit recovery status on every terminal path: repaired, partial or
unrecoverable, plus a byte offset. A candidate that is valid JSON but still fails
envelope validation is marked partial and never returned as a completed result;
the unrecoverable path retains the bounded original payload, the precise parse
error (including its byte offset) and surfaces the status in the FINAL RESULT
VALIDATION diagnostic.
safeToReplay is intentionally unchanged. The deterministic local repair and the
fresh deny-all, no-tools, retained formatter session are independent of the
stage's tool set and never re-prompt or replay the admitted agent session, so a
shell-enabled review stage gains malformed-result recovery without any replay of
admitted work. Risk: no change to mutation safety; the only added cost is one
bounded formatter call, already budgeted. Replayability is not extended for shell
stages.
Docs and packaged skill references (docs/model-results.md,
internal/author/skill/references/{format,debugging,verification}.md) describe the
three-tier recovery, the status words and the derived deadline; the safeToReplay
decision and its risk are recorded there alongside the commit.
Tests add mocked, offline coverage for the deterministic and partial paths, a
repairDeadline monotonicity/floor/clamp table, and a shell-stage regression
proving a recoverable malformed result does not fail the run or replay tool work.