Wrap the judge prompt and two docstrings past the 120-col convention;
record three implementation rulings in the v0.9 spec (exploration read
outlets, round-based sampling, findings carrying all ratings).
report, markdown and analysis each repeated the
summarize_exploration(repo.list_by_campaign(...)) shape; collapse it
into summarize_campaign_exploration so the aggregation has one home.
Ticket 05 asks the judge review conclusions to flow into the report
verbatim; the aggregation silently dropped good/acceptable dimensions.
Collect every finding sorted poor-first and color drawer tags by rating.
E2E on t480 showed assistant bubbles rendering {'content': '...'} because
tutu replies carry msgBody as a parsed object and the router stored
str(reply.content). Coerce to the inner text before persisting.
Exploration sessions aggregate into a single exploration summary
(session counts, goal-achievement rate, issue lists from experience
records, judge conclusions when reviewed) that feeds three exits:
the campaign report gains an exploration dimension, the v0.7 analysis
stage-two input gains the summary (stats only, never full dialogues),
and the Markdown export appends a findings appendix after analysis and
comparison. With no exploration data every output stays unchanged.
After an exploration session closes, the platform samples up to 3
conversation rounds and runs an independent judge-role review through
the v0.7 ChatClient seam, persisting quality-dimension conclusions
(attitude, professionalism, hallucination) into the session's
judge_review. The review runs as a background task: failures are
recorded without touching session state or the first-hand experience
record, and a missing model config skips silently.
Resident agents call GET /api/exploration/patrol once per cycle to see
every running production-line campaign that opted into exploration
(seed set present), the new results since the last watermark (reusing
campaign report aggregation), and the remaining exploration budget.
The watermark advances after each call so subsequent calls only report
increments; accelerated and terminal campaigns are excluded.
v0.9 ticket 02. Campaigns now carry an exploration seed set (seed
personas × seed goals — the comparability unit for exploratory
evaluation) and an optional budget override, stored as JSON columns
isomorphic to plan. Empty seeds normalize to null, marking the campaign
as opted out of exploration. resolve_budget merges per-field overrides
into platform defaults; enforcement stays server-side. The create form
gains seed lists and budget inputs (minutes → seconds), submitting null
when left empty.
v0.9 ticket 01. Independent exploration_sessions/exploration_messages
entities (never merged into EvalRun, keeping ADR-0001/0002 semantics
intact): create/message/close APIs forward virtual-user messages through
the target's real channel, persist both parties' rows with latency, and
close with a whitelist-normalized experience record. Budget enforcement
is a platform ledger — sessions per window, turns per session, and
session interval overruns return 409 with readable reasons; accelerated
lines accept manual sessions only. Messages delivered but unanswered
still consume a turn so timeouts cannot bypass the budget.
The scheduler loop enqueues the analysis task when a realtime campaign
completes; accelerated or cancelled campaigns and a missing analysis
model skip silently. The campaign markdown export appends the analysis
appendix (overall, problems, narratives, suggestions) when a completed
analysis exists.
Add the analysis role's execution path: a two-phase orchestration
(per-scenario diagnosis gathered in parallel, then a synthesis pass)
that reads the existing campaign report aggregation plus capped failure
samples, validates the LLM's JSON against the report schema, and strips
fabricated run/scenario references before persisting. Results upsert one
row per campaign (generating/completed/failed) with the model config
snapshot; GET/POST /api/campaigns/{id}/analysis expose the state machine,
guarding non-terminal campaigns and missing analysis models with 400s.
Campaigns can pin an analysis model config instead of following the
global analysis default. Creation validates the referenced config
exists (400 otherwise); the create form offers enabled chat configs
with the global default as the fallback option.
Introduce ModelPurpose.ANALYSIS and a globally-unique is_analysis_default
marker on chat model configs so campaign analysis can resolve its model.
Service rejects disabled or non-chat configs; repo clears the previous
holder on set. Documented the analysis role in CONTEXT.md.
build_campaign_timeline flattens a campaign's child Runs into offset-sorted
per-Run entries (distinct from the report's 12-bucket aggregation), reusing a
shared _run_window_offset口径 so both views place a run identically. Exposes
GET /campaigns/{id}/timeline and the api.ts type/call. (v0.6 ticket 06)
Campaign.summary was a bare Optional[dict] while RunSummary is a typed VO —
scheduler state (spawned_indices/errors) flowed untyped through
set_/get_summary. Introduce CampaignSummary + SchedulerState (extra=allow,
validate_assignment), mirroring RunSummary; campaign_runner reads/writes the
VO. Also converge the ~10 repeated JSON column get/set pairs onto
_json_dumps/_json_loads helpers, unifying ensure_ascii=False and fixing the
set_modalities ensure_ascii=True trap.
Three near-identical Turn(...)+save_turn blocks (send-fail / poll-except /
happy path) collapse into one _persist_turn helper differing only by the
optional fields set. The期望→隐式规则 translation becomes a pure,
unit-testable derive_implicit_rules seam (CONTEXT: 期望与规则叠加生效),
and the rule_type→ModelPurpose map is hoisted to a module constant.
Single-run summary口径 (pass_rate / judged_pass_rate / avg_latency /
connectivity split) was inlined in run(), reachable only by driving a
whole async run, and report.py recomputed judged_pass_rate independently.
Extract build_run_summary — a pure function parallel to aggregate_runs
(cross-run) and combine_case_outcome (case-level). run() now collects
material and delegates; judged_pass_rate is stored in RunSummary so the
report reads it instead of recomputing.
Read paths recomputed per-case pass/connectivity independently — report
generation, the logs endpoint, and the frontend each derived it, and the
frontend's every(passed) recompute ignored the engine's authoritative
verdict. Extract resolve_case_verdicts: a single pure seam that prefers
stored case_outcomes verbatim and approximates only for legacy runs. The
logs endpoint now surfaces case_verdicts so the frontend reads instead of
recomputing.
report.py mixed DB-reading generation with string formatting: the four
render_*_report(run_id, session) functions each re-fetched via
generate_report, so the HTML/Markdown/JSON formatting was welded to storage
and could not be unit-tested from a plain dict. Extract the formatting into a
new pure report_render module whose renderers take the already-built report
dict (no session, no storage import). Migrate every caller to generate-then-
render, delete the old coupled renderers with no back-compat shim, and drop
the _aggregate_runs middle-man alias in favour of metrics.aggregate_runs.
Both the single-run path and the campaign scheduler drove long-lived
asyncio tasks through their own duplicated _tasks/_cancel_events dicts and
shutdown loops. Collapse them into one deep TaskRegistry module,
instantiated as run_registry and campaign_registry. launch() creates the
cancel event before the task (so a cancel during startup is never lost),
wires done-callback cleanup, and is idempotent per id; this makes runs.py's
hard-cancel fallback provably dead, so it is removed. App shutdown now
gracefully stops in-flight runs too, not just campaigns.
Extend the pure scheduler with elapsed_seconds (clock injected), decide_tick
(offset + due + lifecycle action) and resolve_finalize (cancel-race guard),
so the durable loop stops hand-coding elapsed/finished/status checks and only
does I/O. Deletes the runner's private _elapsed_seconds and converges
current_window_offset onto the one pure elapsed computation. The clock-skew
tolerance and cancel-race guard are now unit-testable at the seam.
Give EvalRun.summary a typed RunSummary value (unified RunError, lenient
legacy parsing) so readers stop reaching into a schemaless dict, and route
every cross-run rollup — dashboard, scenario ranking, trend, campaign
report — through one aggregate_runs seam. Fixes the divergence where
stats averaged pass_rate over completed-only runs while the campaign
report counted faults as 0.0. Cross-run rule (ADR-0004): genuine faults
count 0.0, user-cancelled runs are excluded from both denominators.
Embed compact progress (completed/planned total + overall pass_rate,
reusing the report's aggregation) into GET /campaigns so the list drops
its N+1 report fetch. Poll list and open report drawer every 5s while the
tab is active and a campaign is still running. Show scenario version and
trigger source tags in the child-run drill-down.
Add generate_campaign_report: a pure aggregator over a campaign's child Runs
producing a time-trend axis (Runs bucketed by service-window position) and a
capability-summary axis (grouped by scenario), each carrying pass_rate /
availability / latency. pass_rate keeps the single-Run case-level meaning and
counts execution failures as 0.0 (ADR-0002); time_scale only places Runs into
window-time buckets and never alters any figure. Engine summary now records
avg_latency_ms to feed the latency axis.
Expose GET /api/campaigns/{id}/report (structured) and .../report/markdown
(reusing the existing Markdown export path). Adds "可用性/Availability" to the
domain glossary.
Add a thin async loop (run_campaign_loop) that ticks on real wall-clock time,
maps elapsed×time_scale to a window offset via the pure decide_schedule, spawns
due child Runs, and marks the campaign COMPLETED at window end. All authority
lives in the DB (started_at, spawned_indices, status), so the app lifespan can
resume every RUNNING campaign on startup without double-spawning and stop all
loops gracefully on shutdown. A failing plan entry is skipped and recorded
rather than wedging the campaign.
Creating a campaign now starts its loop; POST /api/campaigns/{id}/cancel stops
further spawning (completed child Runs are kept); GET /api/campaigns/{id}
reports live progress (window offset, spawned/completed Run counts).
Add the pure scheduling seam (campaign_scheduler.decide_schedule) that, given a
static plan and window-clock offset, decides which plan entries are due and
whether the window ended — mirroring judgement.combine_case_outcome, with
time_scale confined to the clock mapping so it never touches judgement/report.
The campaign_runner shell maps injected elapsed time to a window offset, spawns
due child Runs through the existing EvalEngine.run(existing_run=...) path with
campaign_id + RunTrigger.CAMPAIGN, and persists spawned-entry indices per entry
for idempotent, restart-recoverable progress. No auto loop yet (ticket 03).
