Participant: codex (AI agent)
Understanding
Unbounded repository-wide LLM passes waste context on low-risk files, while a
deterministic-only analysis misses errors of intent. The useful boundary is a
deterministic policy that chooses semantic stages and evidence slices, followed
by required LLM reasoning inside every selected semantic stage. Token/request
ceilings and exact cache identity make cost auditable without silently changing
the analysis method.
Execution plan
- Commit this plan and intent before any implementation file.
- Define a closed policy vocabulary, typed representation and canonical text
codec in the core-dsl workstream.
- Validate stage topology and nested budgets fail closed.
- Add deterministic selection, usage aggregation and exact policy
fingerprinting with focused tests.
- Attempt a real todo2code
require-llm analysis, then run host, governance,
dependency and Docker gates.
Actual changes
- Plan and intent committed independently as
f2e68ae; implementation has not
started before the separately recorded approval transition 84690c1.
- Added the dependency-free
t2c.analysis-policy/v1 contract with a canonical
parser/renderer, closed vocabulary, fail-closed validation, deterministic
stage selection, usage aggregation, policy/cache fingerprints and a
provider-rate cost estimator.
- Made global and per-stage budgets bound requests (including retries and
response repair), input tokens, output tokens and elapsed time. Selected
semantic stages require LLM; provider failure never becomes a deterministic
semantic result.
- Added eight focused tests covering byte identity, selection, exact evidence
caching, cost bounds, invalid vocabulary/topology/budgets and malformed text.
- Ran the real todo2code pipeline with all semantic modes set to
require-llm.
It issued 43 provider responses, consumed 498,533 input and 153,176 output
tokens and cost 1.303208672 USD. The 920,279 ms Markdown stage caused the
contract to gain explicit elapsed-time ceilings.
- Applied the LLM’s ticket-local wording finding by replacing ambiguous nominal
constraints with explicit operational verbs. Historical findings lacking
evidence in the owned paths were rejected.
- Final local validation: focused 8/8; full host 414 pass, 0 fail, 1 JDK skip;
gold gates 100%; governance 0 findings; audit 0 vulnerabilities; Docker
smoke PASS.
- Protected PR #86 passed Koru, hosted checks and review-triggered governance.
Validator-agent run
31580896098 approved all four exact-head chunks with
GLM-5.2 and no findings. Publication merged as main@f0df52c; post-merge CI
31581196362 passed every required job.
Blockers
- None. Ticket and workflow are complete.