Capability gate: promptfoo codex-sdk harness for the Codex jcodemunch, ai-memory and M13 gates, reconciled with native OTel - #433

… gate rows

The decision addendum and the evidence README said #433 was open. Its merged body records the first smoke's route miss (6 of 6), the repaired-scoring rows (order search_symbols 6 of 6, each returning the signature) and the pre-repair smokes as superseded; the two passages now say so.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…echeck stamps, stack revision (#437 review H5a-d)

(a) Summaries run to 606 characters, not 600: summarize() cuts at 600 and
appends the 6-character " [...]" marker; 138 of 206 are cut and 60 exceed 600.
The subset's own method field keeps the generator's wording and stays as
generated. (b) The 38 exit-null records are 12 MCP requests (seven Claude Code
tool calls: six socraticode, one qmd; five mcporter requests to repomix) and 26
shell plumbing, upstream CLI runs, a capture harness and local checks, not
"mostly MCP calls". (c) The socraticode, qmd and ai-memory release rechecks are
stamped 2026-09-27T01:11:26Z-01:24:04Z (2026-09-26 local). (d) condense.py stops
at the first card whose pin differs from manifests/stack.json; the cards assume
the stack manifest at 34c5634 (#372), which the rtk and toon cards cite, and
55fc8d1 (#433) has the same pins. The Reproduce section runs condense.py in a
temporary root with that manifest; the documented block exited 0 with identical
bytes at 2026-09-29T03:28Z (also with 55fc8d1). It also restores the pin
sentence the previous commit split. Tests derive each statement from the
bundle's data and fail on the previous README.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

… drift notes and a weekly native E2E (#437)

* Token topic: 2026-09-27 per-tool card edition, stack-pinned join and drift note

docs/token-efficiency-stack.json gains a dated edition. 17 rows carry a card
condensed from the sanitized token-stack cards: upstream, native adaptation,
E2E returned results, adapted performance per payload and lane, invoke rates
and GPT-6 review, each labelled with its evidence class. The other 7 rows
carry an explicit "no card in this edition" marker. gpt-tokenizer has a card
but no stack.json component, so it is listed under edition.cards_without_row.

scripts/build_ecosystem.py joins each row to manifests/stack.json for the
version and repository and rejects row-level pins. When a card's recorded pin
differs from the stack pin, the row shows a drift note instead of editing the
card. The build validates every card block's evidence class, each
comparison's counts and the card's source bytes and sha256. The edition and
its artifacts link at the publication ref.

evidence/artifacts/token-stack-cards-20260927 holds the sanitized cards, the
invoke snapshot, the 206 cited returned-result records (command, exit,
summary and the sha256 of each omitted body), the GPT-6 review index, a README
with evidence classes and dates, and the generators.

native-token-e2e.yml adds a weekly Monday 07:11 UTC schedule. AGENTS.md and
docs/token-session-handbook.md each gain one pointer to the rows.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

* Token topic page: render the per-tool cards (template outside the brief's path list — coordinator to accept or reassign)

The page's only inline script lives in docs/ecosystem/template.html, so the
edition header, the stack-pinned drift note, the no-card marker and the six
card blocks (upstream, native adaptation, E2E returned results, performance
per payload and lane, invoke rates, GPT-6 review) render there. Each block
states its evidence class; figures keep their lane, payload and unit and are
never added across them. Only existing helpers (make/add/link with safeHref,
textContent) are used; the build recomputes the CSP script hash.

tests/test_ecosystem_manifest.py runs the real template through the Node
page harness and asserts the rendered card text, the drift note and the
no-card marker.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

* Hot files: evidence registration and generated reports

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

* Token topic cards: card link fields pass the build's public_url gate (#437 review H2)

docs/ecosystem/template.html topicCard renders upstream.recommended_install.url,
recommended_wiring[].url, new_since_pin[].url and limitations[].url as links, but
token_topic_card returned the card unchanged, so these reached safeHref, which
accepts any http/https URL. Each now passes public_url, the gate card.source and
the build's other external source links use; a rejected URL fails --check with
the card's source path and the field. Tests plant http://, 127.0.0.1 and
javascript: URLs in each field (12 cases, all passed the unchanged build) and
accept a clean HTTPS set unchanged.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Token stack cards: bundle.py keys the returned results by the supplied basename (#437 review H5e)

source() records each input under clean_text(path.name), but main() read
inputs["returned-results.json"], so renaming the --returned-results file raised
KeyError at line 190. The subset's source block now uses the supplied basename
for both the lookup and its name; with the canonical name the output is
unchanged. tests/test_token_stack_cards.py runs bundle.main on a synthetic
input set under both names (the renamed case raised KeyError before the fix).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Token stack cards: retained ledger readings behind the native_snapshot figures (#437 review H1)

The five native_snapshot figures (rtk 42,367,855 and 26,531,271, headroom
579,411, context-mode 3,244,032 and 6,311,424) cited private assembly inputs
only. tools/counter_readings.py opens the private token-report ledger read-only
(sqlite3 URI mode=ro), matches each entry to its one snapshots row on tool,
$HOME-relative scope, observed_at to the second and saved figure, requires the
kind and session estimate the entry's basis quotes to equal the row's, and
writes counter-readings.json: ledger ids 67, 68, 69, 71 and 72 with tool, scope,
observed_at, success, saved, kind and, for context-mode, the session estimate
(27,109,948 and 157,255). The evidence column and metrics.raw are never copied;
the output passes a private-content scan before it is written. The copy re-measures nothing.

The README's new section tabulates the backed figures and names the two rtk
shortfall figures the card cites without a reading (39.2% over the host's
lifetime, 16,493 rtk discover skips): they rest only on the private rtk.json
input (sha256 in the Reproduce table) and this artifact does not verify them.
The jcodemunch-mcp card has no native_snapshot entry. Tests require an
identical reading for every entry, fail when one reading's saved figure, scope
or time changes, allow only the sanitized fields and check the README section.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Token stack cards README: measured summary bound, exit-null counts, recheck stamps, stack revision (#437 review H5a-d)

(a) Summaries run to 606 characters, not 600: summarize() cuts at 600 and
appends the 6-character " [...]" marker; 138 of 206 are cut and 60 exceed 600.
The subset's own method field keeps the generator's wording and stays as
generated. (b) The 38 exit-null records are 12 MCP requests (seven Claude Code
tool calls: six socraticode, one qmd; five mcporter requests to repomix) and 26
shell plumbing, upstream CLI runs, a capture harness and local checks, not
"mostly MCP calls". (c) The socraticode, qmd and ai-memory release rechecks are
stamped 2026-09-27T01:11:26Z-01:24:04Z (2026-09-26 local). (d) condense.py stops
at the first card whose pin differs from manifests/stack.json; the cards assume
the stack manifest at 34c5634 (#372), which the rtk and toon cards cite, and
55fc8d1 (#433) has the same pins. The Reproduce section runs condense.py in a
temporary root with that manifest; the documented block exited 0 with identical
bytes at 2026-09-29T03:28Z (also with 55fc8d1). It also restores the pin
sentence the previous commit split. Tests derive each statement from the
bundle's data and fail on the previous README.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Token session handbook: token topic rows are the adoption reference, cards are dated snapshots (#437 review H4)

The line this PR added made each tool's topic row "its lane and adoption
reference", but the cards are dated 2026-09-26/27 snapshots and the qmd card's
lane_rule is already superseded by the SubagentStart carrier
(adoption/hooks/claude/token-lanes-block.md L9 also names foundation-docs and
foundation-adoption). The row is now the adoption reference, the current lane
list is the carrier block, and each card block is called a dated snapshot.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Hot files: AGENTS.md token-topic pointer and evidence registration (#437 review H4)

AGENTS.md: the pointer this PR added now makes a tool's topic row its adoption
reference, calls its card a dated snapshot, and names the SubagentStart carrier
block (adoption/hooks/claude/token-lanes-block.md) as the current lane list; one
line, no other line touched.

manifests/evidence.json registers the changed listed files (AGENTS.md,
docs/token-session-handbook.md, scripts/build_ecosystem.py,
tests/test_ecosystem_manifest.py, the card bundle README and tools/bundle.py)
and the new evidence files counter-readings.json and tools/counter_readings.py.
component_matrix.py --write and new_host_grand_list.py --write left their four
reports unchanged.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* Token stack cards: reproduction block status, truncation condition, list guard, test comment (#437 verifier findings)

The round-one verifier found the README's reproduction block reported cmp's status, so it passed even when condense.py stopped: the block now runs under set -e and the README states the measured exit statuses of the block as written and with the current stack manifest. The truncation sentence names its condition (cut at the last space only when it is past character 300). token_topic_card's list guard now rejects any non-list, including a falsy one: a test covers {}, "" and null and fails against the previous guard. A test comment states where card.source.url comes from.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

* AgentsView re-pin registry: count the per-tool card lines that name the pinned version

#437 adds per-tool card blocks to docs/token-efficiency-stack.json. The agentsview
card names 0.43.0 on 12 lines as dated card facts (the pin the card recorded), so
main's RepinLocationRegistryTests, which pairs every matching line of a re-pin
file with a recorded location, failed with 14 matching lines against 2 recorded
locations (CI run 36526587362). The registry entry now records dated_record_lines
(count 12, with its basis) and location_mismatches counts them exactly; a wrong
count fails a new test. A re-pin still changes only the two install lines.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

* Hot files: re-register the AgentsView registry files in the evidence manifest (manifests/evidence.json only; last commit per docs/lanes.md)

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

---------

Co-authored-by: Scout <scout@local>
Co-authored-by: Claude Opus 5.5 (1M context) <[email protected]>