Plan execution feedback

Plan Execution Feedback Skill

After each session, compares planned vs actual (write-time fixes, test counts, API assumptions). Tracks discovery rate over time. When discovery rate consistently exceeds 120%, flags the planning process as under-investing in G0 infrastructure discovery.

When to Invoke

• At session end, after Section A is written
• Before writing the next session's plan (to calibrate G0 depth)

Metrics

WF discovery rate = (WF actual) / (WF planned) × 100%
Test delivery rate = (tests delivered) / (tests planned) × 100%

Interpretation Table

WF Discovery Rate	Signal	Required Action for Next Plan
≤ 120%	Normal	Proceed with standard G0
121–200%	G0 under-invested	Add explicit api-shape-verification table to next plan
> 200%	API assumptions are guesses	Do NOT write plan without live source reads first

XIIGen Execution History

Session	WF Planned	WF Actual	Discovery Rate	Signal
S1	0	0	—	Normal
S2	0	5	∞%	G0 missed API shapes entirely
S3	4	7	175%	G0 under-invested
S4	7	?	TBD	Must complete api-shape-verification table at G0

Pattern: Sessions 2 and 3 both had high discovery rates. This means the planning process consistently underestimates API mismatches. Root cause: plan authors write tests against assumed APIs without reading source files first.

Rules

1. Record WF planned vs actual in STATE.json after every session
2. If discovery rate > 120% for two consecutive sessions: the next session plan MUST include a completed api-shape-verification table before any code blocks are written
3. If test delivery rate < 80%: investigate whether the plan's test count was based on counting describe blocks vs it() blocks (common source of over-counting)
4. Report the discovery rate metric as part of Section A: WF planned: N, WF actual: M, discovery rate: M/N×100%

Integration

• Invoked by agent-constitution at Session End (after Section A)
• Feeds into planning-skill G0 depth calibration for the next session
• Results stored in STATE.json under phases.N.wf_discovery_rate

---

UNIVERSAL STANDARD ADDENDUM — Human-Readable Log Gate + User Prompt Source update (ported from llm_mvp_core)

Added by the Universal-Skills refresh (UpdateUniversalSkills). The 120%/200% thresholds above are kept. The universal layer adds a readability gate on the feedback record and a per-phase update to the User Prompt Source Ledger, so the metric is never machine-only.

Human-Readable Plan/Execution Log Gate (Section A is not just machine rows)

A per-phase feedback record FAILS this gate if a human cannot tell from it: what the phase did, what is done, what remains, which claims stayed closed and why, and the next bounded action. Numbers (discovery rate, test delivery rate) are secondary evidence — they appear AFTER a plain-language line.

Phase N of M: <human-readable name>. Done: <…>. Remaining: <…>.
Discovery rate: WF planned N, WF actual M, M/N×100%.   # machine row AFTER the human line
Test delivery: delivered/planned ×100%.
Next action: <next bounded step, or NONE>.

A discovery rate / test delivery row with no preceding human sentence is a machine-only record and does not satisfy Section A.

User Prompt Source Ledger — update each phase

After each phase, append to the User Prompt Source Ledger any requirement that was discovered or strengthened at write-time (a wrong assumption about a NestJS service shape, a missing React/FastAPI contract, a DTO mismatch). A write-time fix that revealed a new requirement is recorded as its source so the next plan's Gate-0 discovery covers it. This closes the loop between plan-execution-feedback and planning-session-startup (Authority/Source/Delta ledgers).

Discovery-rate over-investment signal (>120% restated as a learning signal)

WF discovery rate > 120% for two consecutive phases is not just "add an api-shape-verification table". It is evidence that Gate-0 discovery under-read live TS source (server/src/**.ts, server/src/engine-contracts/*.ts). The next plan MUST read those sources live before writing test code (Jest/Playwright), and the feedback record must say so in plain language.