Planner becomes the sole case-SDD author - #2431
Draft
abhiram-vad wants to merge 10 commits into
Draft
Conversation
…tion (1/4) Planner becomes the sole case-SDD author. Adds the conversational case design lane (Listen/Sketch/one Case Review confirmation) with three entry modes (delegated SDD-only from uipath-maestro-case, direct design, draft finalization), full design-time tenant resolution (registry pull, cache lookups, connections, ambiguity/empty gate at review time), the case authoring rules (ported from maestro-case sdd-generation-rules), worked I/O examples, and the blueprint-shaped case SDD template carried forward from #2305 with the Testing/NFR un-nesting fix. Delegated mode renders SDD text in memory and hands back — the caller owns the write. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… 1 (2/4) When sdd.md is absent, entry delegates the design conversation to uipath-planner (SDD-only mode) and lands the returned model in ONE batched write (sdd.md + tasks/tasks.md + registry-resolved.json + solution init). Degraded path: planner absent → ask for sdd.md and stop. phase-0-interview.md is deleted; the draft-finalization carve-out and 'never auto-invoke uipath-planner' rule are retired. Phase 1 becomes verify-only against the planner's resolution ledger (re-resolve only on staleness/cross-session); the Rule-17 gate short-circuits on recorded design-time gate decisions. Rides along: validate-loop guard (never re-validate without an intervening edit) and the in-session schema memo (never re-run an identical tasks describe / case spec). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Design-side tasks move to the planner suite: the two phase-0 interview evals (candidate_interview, loan_origination) and the four finalize-from-draft variants. procurement_sla_interrupts stays in maestro-case renamed design_delegation_procurement — it exercises the runtime delegation (DAG 1). New build_from_planner_sdd task proves a Planner-Handoff-bearing sdd.md does not hijack maestro-case into Lane A task derivation. Moved checkers bootstrap sys.path against the maestro-case _shared dir; _shared unit tests parameterize checker paths. Local verification (Bedrock, single replicates): migrated finalize_draft 1.0 SUCCESS 8/8 vs baseline main 0.783 FAILURE (backticked <UNRESOLVED> broke the name-contract checker); build_from_planner_sdd 1.0 SUCCESS 7/7. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Skill descriptions changed in commits 1-2 (both skills — expect the double activation-gate recall eval); CODEOWNERS follows the moved test paths. skill-status.json unchanged: both skills keep their status. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Transcript-mined fixes from the migration eval runs. The 50-min finalize run decomposed as: 17.5 min composing the whole SDD in-head before any Write (destroyed by auto-compaction), ~20 min post-compaction flailing (skill re-invocation, filesystem hunts, orphaned background tasks), and a 1015-line authoring-guide read the finalize path never needed. - Case SDD writes are now write-early + section-batched: seed with Status: draft, per-section Edit-appends, conformance gate on the on-disk file, ready flip last. Compaction recovery = resume from the partial file; never re-invoke skills or re-read applied guides. - Hard finalize read budget: lane-guide finalize sections + draft + template, once each; authoring-rules guide and examples explicitly excluded. Delegation happens before any maestro-case reference reads. - Plain <UNRESOLVED> markers: template placeholders de-backticked and the rule stated in the gate — backticked markers broke the name-contract checker on both the baseline and one fixed-run attempt. - Build phases: read each plugin planning.md/impl-json.md once per plugin type, not per T-entry (observed 10-26x re-reads per build). Measured on the same finalize eval (Bedrock, single replicates): pre-fix 2995s/114 turns SUCCESS 1.0 -> post-fix 527s/12 turns SUCCESS 1.0 (5.7x), vs baseline main 1106s/63 turns FAILURE 0.783. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…gates sdd-generation-rules.md (maestro-case) drops the 10 authoring-only sections the planner now owns (task-type choosing + override priority, authority hierarchy, source ledger, domain fidelity, architect's lens, finalization checklist, worked reasoning, inputs, anti-patterns): 1015 -> 774 lines. What remains is the build-side parse contract the plugins and validators cite (content rules, outputs grammar, lineage, logical integrity, review-item shape, I/O completeness) — all inbound anchors verified. Dangling references to cut sections rewritten. Planner: the authoring guide's duplicated template-gate shape compresses to a pointer at the lane guide's single source; a read map at the top scopes reads per activity (finalization reads none of the file). maestro-case: stale 'Phase 0' labels renamed to design delegation across SKILL.md, planning.md, brownfield.md, registry-discovery.md, and the SLA plugin; the Case Review section list is enumerated once (Rule 1) instead of three times. Re-verified after the cuts (Bedrock, single replicates): finalize_draft SUCCESS 0.99 (543s/11 turns), build_from_planner_sdd SUCCESS 1.0 (161s/11 turns). Link+anchor sweep clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…shared The migrated planner design evals were still reaching into the maestro-case test folder: YAML graders called tests/tasks/uipath-maestro-case/_shared/sdd_check.py, four checkers sys.path-bootstrapped across suites, and the maestro-case _shared unit test file reached back into planner task dirs. Per-folder ownership now holds: - tests/tasks/uipath-planner/_shared/ created: entry_rule_check.py MOVES (only planner checkers use it); case_check.py / sdd_check.py / test_sdd_check.py are planner-owned copies (both suites use them); test_finalize_checkers.py carries the picker/reject/parser behavioral matrix split out of test_reentry_reachability_checks.py. - maestro-case _shared keeps only what its own suite uses; its unit test file retains the caseplan-emit JMESPath matrix. - All planner checkers and YAML grader paths repointed inside their suite; CODEOWNERS row added for the new _shared. Verified: 42 + 52 unit tests pass per suite; planner sdd_check copy re-graded a real run artifact identically; both changed YAMLs pass TaskDefinition schema; cross-folder grep clean (comments only). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…versight carve-out ambiguity Found by a live E2E run of the migrated flow (loan origination, published to Studio Web): the agent paired return-to-origin with Interrupting: No on an SLA parallel-oversight lane. The carved-out oversight exception sentence read as if it also relaxed the returning-lane rule; CLI validate is blind to the pair (passed Valid), only the sdd_check grader catches it. Both copies (planner authoring guide + maestro-case build contract) now state: oversight lanes complete exit-only; return-to-origin always requires Interrupting: Yes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…undary Design decision (session, 2026-08-03): replace the in-conversation SDD-only hand-back with a simpler, durable contract. uipath-maestro-case spawns uipath-planner as ONE subagent; the subagent runs the Case Design Lane best-assumption (no user reachable), does full design-time resolution, WRITES sdd.md at the working root via the write-early cadence, and returns the Case Review packet + decisions + resolution ledger (fenced JSON) in its final report. maestro-case then presents the returned review to the real user as the single confirmation (Rule 11 build options), reads sdd.md once as sole input, and builds; corrections re-delegate as targeted sdd.md edits. Planner now ALWAYS writes the SDD in every mode; the model no longer crosses the skill boundary — the file is the artifact at the seam (compaction-proof), the review is the approval payload, the ledger still powers verify-only Phase 1 (the ~/.uip cache is machine-global, so the subagent's pull is the session's pull). Degraded path unchanged. Cross-skill eval gains an allowed_tools override (Agent/Task) so the subagent can spawn under coder-eval. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…g files Per review direction: the migration is a clear lift and shift, not a copy. Removed from uipath-maestro-case: sdd-generation-rules.md (the authoring/content contract — planner's case-authoring-rules-guide.md is now the single home), sdd-template.md, sdd-template-examples.md; moved sdd-viewer.html to planner (the SDD author owns the review surface, lane guide § HTML preview restored). All 11 inbound reference sites repointed: build-load-bearing links now target build-side homes (io-binding Check 5, io-binding planning projection grammar, global-vars impl-json category semantics); design-side concepts (logical-integrity divert contract, Section 4 roll-up, SLA response model) become textual mentions of the planner-owned contract — no cross-skill file links, per self-containment. tests/pytest.ini adds importlib import mode so the two per-suite _shared packages collect in one run; the moved parser test imports its checker by file path. 94/94 unit tests pass; link sweep and description/status hooks green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
TL;DR
Phase 0 case-design authoring moves from
uipath-maestro-casetouipath-planner. maestro-case delegates the design conversation at runtime (same context window — the in-memory model crosses the skill boundary for free), consumes the confirmed model, and landssdd.md + tasks/tasks.md + registry-resolved.json + solution initin ONE batched write. Planner does FULL tenant resolution at design time (registry pull, per-resource cache lookups, connections, ambiguity/empty gate at Case Review), so Phase 1 becomes verify-only. Subsumes and closes #2305 (its case-sdd-template + guide content is carried forward in commit 1, plus the Testing/NFR un-nesting fix).Decisions (locked 2026-07-30 design review)
What changed
case-design-lane-guide.md(delegated SDD-only / direct / draft / finalization entries, design-time resolution gate, template gate, terminal-step fork), NEWcase-authoring-rules-guide.md(ported authoring rules), NEWcase-sdd-examples.md, Entry Guard step 0 (delegation detect), Rule 13phase-0-interview.mddeleted; Rule 1 = delegation + ONE batched write; Rule 15 flipped (never-auto-invoke → delegation contract + degraded path: planner absent → ask for sdd.md & stop); Rule 3 verify-only w/ planner ledger; Rule 17 short-circuit on recordedgateDecision; validate-loop guard + in-session schema memo inphased-execution.mdtests/tasks/uipath-planner/;procurement_sla_interrupts→design_delegation_procurement(cross-skill DAG-1 eval); NEWbuild_from_planner_sdd(Planner-Handoff sdd.md must not hijack into Lane A)Verification (local coder-eval, Bedrock, single replicates)
case_finalize_draft(planner-authored)main<UNRESOLVED>, breaking the name-contract checkbuild_from_planner_sdd_sharedunit testsNot yet run: interactive DAG-1 live-tenant E2E (50–90 min class), interview design evals + 3 finalize variants (same code paths as the verified one), activation-gate recall (CI — both descriptions changed, double gate expected).
Closes #2305
🤖 Generated with Claude Code
Perf commit (added after transcript mining)
The migration eval transcripts exposed three latency sinks; commit 5 fixes them at doc level and re-measures on the same finalize eval:
main<UNRESOLVED>broke the name-contract check)Fixes: write-early section-batched SDD render (seed
Status: draft→ per-section appends → gate on-disk → ready flip last, with a compaction-recovery protocol), hard finalize read budget (lane-guide finalize sections + draft + template only), plain<UNRESOLVED>markers (template de-backticked — this defect also failed the baseline run), immediate delegation before any maestro-case reference reads, and plugin read-once for build phases (observed 10–26×planning.md/impl-json.mdre-reads per build).