Skip to content

docs: authorize bounded P1 v8 retry #299

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ictechgy merged 1 commit into from
Aug 11, 2026
Merged

docs: authorize bounded P1 v8 retry #299

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
42 changes: 30 additions & 12 deletions research/p1-live-authorization-packet.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,8 @@

_Drafted: 2026-08-09 KST; amended with explicit auth reuse, stopped-root
checkpoints, the exact 264-identity v6 ceiling approval on 2026-08-10 KST,
and the exact 307-identity v7 ceiling approval on 2026-08-11 KST_
and the exact 307-identity v7 and 388-identity v8 ceiling approvals on
2026-08-11 KST_

This packet is the controlling scope for the authorization recorded on
2026-08-09. Approved actions are repository-scoped GitHub activity, candidate-
Expand Down Expand Up @@ -107,6 +108,23 @@ this packet.
explicit cumulative ceiling of at least 388. Provider-free attempt-v4 TDD
repairs the bounded-failure classification for a future reviewed candidate,
but does not alter, migrate, resume, or reinterpret the v7 P1-X record.
- On 2026-08-11 the user approved one fresh v8 root under the unchanged finite
study shape below, raising the cumulative consumed/reserved ceiling from 307
to exactly 388. The grant retains all 170 terminal identities and authorizes
at most 218 new identities: two discarded canaries, 108 initials, and up to
108 policy-valid retries. The model remains `sonnet`, every process remains
capped at `$0.75`, and the fresh-root arithmetic maximum remains `$163.50`.
Old-root reuse, replay, repair, migration, optional stopping, npm publication,
and active P2-P6 work before P1-F remain forbidden. Integrity, privacy,
ambiguity, or spend failure must immediately close the root to provider work
and produce the canonical claim-disabled P1-X.
Comment on lines +117 to +120

@coderabbitai coderabbitai Bot Aug 11, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🗄️ Data Integrity & Integration | 🟠 Major | 🏗️ Heavy lift

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
set -euo pipefail

rg -n -C 8 \
  'P1-X|claim_allowed|stop_reason|integrity|privacy|ambigu|spend|reservation|cancel|drift|HOME|credential' \
  --glob '*.py' .

Repository: ictechgy/context-guard

Length of output: 50379

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '%s\n' '--- contract ---'
sed -n '100,135p' research/p1-live-authorization-packet.md

printf '%s\n' '--- v2 symbols and P1-X references ---'
rg -n \
  'BENCHMARK_STUDY_V2|P1-X|invalid_.*decision|decision|stop_reason|claim_allowed|raise (ValueError|RuntimeError|SystemExit)|spend|privacy|integrity|credential|reservation|cancel|drift' \
  context-guard-kit/benchmark_runner.py \
  --max-columns 220 \
  | head -n 1600

printf '%s\n' '--- function outline ---'
ast-grep outline context-guard-kit/benchmark_runner.py | sed -n '1,260p'

Repository: ictechgy/context-guard

Length of output: 50378

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '%s\n' '--- decision builder and persistence references ---'
rg -n \
  '_benchmark_study_v2_invalid_|invalid_decision|P1-X|decision_path|decision.json|write.*decision|json.*decision' \
  context-guard-kit/benchmark_runner.py \
  --max-columns 240

printf '%s\n' '--- v2 lifecycle sections ---'
sed -n '13280,14380p' context-guard-kit/benchmark_runner.py

Repository: ictechgy/context-guard

Length of output: 50378

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '%s\n' '--- v2 dispatch and decision writing ---'
sed -n '14370,14580p' context-guard-kit/benchmark_runner.py

printf '%s\n' '--- v2 spend and stop-related branches ---'
rg -n -C 5 \
  'study-v2|BENCHMARK_STUDY_V2|budget|cost_usd|max_budget|reservation|launch_reserved|blocked_study_invalid|infrastructure-invalid|integrity|privacy|credential|HOME|cancel' \
  context-guard-kit/benchmark_runner.py \
  --max-columns 240 \
  | awk 'NR <= 1800'

printf '%s\n' '--- v2 tests mentioning invalid decisions and stop behavior ---'
rg -n -C 8 \
  'invalid decision|invalid_decision|P1-X|blocked_study_invalid|ambiguous.*process|infrastructure-invalid|claim_allowed|spend|budget' \
  tests context-guard-kit --glob '*test*.py' \
  --max-columns 240 \
  | head -n 1600

Repository: ictechgy/context-guard

Length of output: 50379

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '%s\n' '--- exact v2 action handler ---'
rg -n -C 20 \
  'args\.study_v2_action|execute_benchmark_study_v2|analyze_benchmark_study_v2_executable|study-invalid-decision\.json|report_name' \
  context-guard-kit/benchmark_runner.py \
  --max-columns 240

printf '%s\n' '--- exact decision output branch ---'
sed -n '14380,14470p' context-guard-kit/benchmark_runner.py

Repository: ictechgy/context-guard

Length of output: 21537

🏁 Script executed:

#!/bin/bash
set -euo pipefail

python3 - <<'PY'
import ast
from pathlib import Path

path = Path("context-guard-kit/benchmark_runner.py")
tree = ast.parse(path.read_text(encoding="utf-8"))

wanted = {
    "_execute_benchmark_study_v2_canary_unlocked",
    "_execute_benchmark_study_v2_unlocked",
    "_benchmark_study_v2_run_slot",
    "_benchmark_study_v2_invalid_canary_decision",
    "_benchmark_study_v2_invalid_analytic_decision",
    "_analyze_benchmark_study_v2_executable_unlocked",
    "main",
}

def text(node):
    return ast.get_source_segment(path.read_text(encoding="utf-8"), node) or ""

for node in ast.walk(tree):
    if not isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)) or node.name not in wanted:
        continue
    calls = []
    raises = []
    literals = []
    for child in ast.walk(node):
        if isinstance(child, ast.Call):
            fn = child.func
            if isinstance(fn, ast.Name):
                calls.append(fn.id)
            elif isinstance(fn, ast.Attribute):
                calls.append(fn.attr)
        elif isinstance(child, ast.Raise):
            raises.append((child.lineno, text(child)))
        elif isinstance(child, ast.Constant) and isinstance(child.value, str):
            if child.value in {
                "P1-X", "error_max_budget_usd", "valid_task_failure_v1",
                "study_infra_invalid", "ambiguous provider process state permanently blocks this study root",
                "terminal infrastructure-invalid evidence permanently blocks later provider launches",
                "v2 {arm} host PreToolUse canary failed",
            }:
                literals.append((child.lineno, child.value))
    print(f"\nFUNCTION {node.name} line {node.lineno}")
    print("CALLS", sorted(set(calls)))
    print("RAISES", raises[:12])
    print("KEY_LITERALS", literals)

print("\nP1-X persistence sites:")
for node in ast.walk(tree):
    if isinstance(node, ast.Call) and isinstance(node.func, ast.Name) and node.func.id == "_study_write_private":
        source = text(node)
        if "report" in source or "decision" in source:
            print(node.lineno, source.replace("\n", " "))

print("\nBudget classification source hits:")
source = path.read_text(encoding="utf-8").splitlines()
for i, line in enumerate(source, 1):
    if "BENCHMARK_STUDY_V2_BOUNDED_FAILURE_RESULT_CODES" in line or "bounded_failure_usage" in line or "valid_task_failure_v1" in line:
        if 13600 <= i <= 13920:
            print(i, line.strip())
PY

Repository: ictechgy/context-guard

Length of output: 7387

🏁 Script executed:

#!/bin/bash
set -euo pipefail

python3 - <<'PY'
import ast
from pathlib import Path

path = Path("context-guard-kit/benchmark_runner.py")
source = path.read_text(encoding="utf-8")
tree = ast.parse(source)

wanted = {
    "_execute_benchmark_study_v2_canary_unlocked",
    "_execute_benchmark_study_v2_unlocked",
    "_benchmark_study_v2_run_slot",
    "_benchmark_study_v2_invalid_canary_decision",
    "_benchmark_study_v2_invalid_analytic_decision",
    "_analyze_benchmark_study_v2_executable_unlocked",
    "main",
}

def segment(node):
    return ast.get_source_segment(source, node) or ""

for node in ast.walk(tree):
    if not isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
        continue
    if node.name not in wanted:
        continue

    calls = []
    raises = []
    literals = []
    for child in ast.walk(node):
        if isinstance(child, ast.Call):
            if isinstance(child.func, ast.Name):
                calls.append(child.func.id)
            elif isinstance(child.func, ast.Attribute):
                calls.append(child.func.attr)
        elif isinstance(child, ast.Raise):
            raises.append((child.lineno, segment(child)))
        elif isinstance(child, ast.Constant) and isinstance(child.value, str):
            if child.value in {
                "P1-X",
                "error_max_budget_usd",
                "valid_task_failure_v1",
                "study_infra_invalid",
                "ambiguous provider process state permanently blocks this study root",
                "terminal infrastructure-invalid evidence permanently blocks later provider launches",
                "v2 {arm} host PreToolUse canary failed",
            }:
                literals.append((child.lineno, child.value))

    print(f"\nFUNCTION {node.name} line {node.lineno}")
    print("CALLS", sorted(set(calls)))
    print("RAISES", raises[:12])
    print("KEY_LITERALS", literals)

print("\nP1-X persistence sites:")
for node in ast.walk(tree):
    if isinstance(node, ast.Call) and isinstance(node.func, ast.Name):
        if node.func.id == "_study_write_private":
            text = segment(node).replace("\n", " ")
            if "report" in text or "decision" in text:
                print(node.lineno, text)

print("\nBudget classification source hits:")
for i, line in enumerate(source.splitlines(), 1):
    if (
        13600 <= i <= 13920
        and (
            "bounded_failure_usage" in line
            or "valid_task_failure_v1" in line
            or "BENCHMARK_STUDY_V2_BOUNDED_FAILURE_RESULT_CODES" in line
        )
    ):
        print(i, line.strip())
PY

Repository: ictechgy/context-guard

Length of output: 7387

Route every v8 stop path through persisted P1-X.

study-invalid-decision.json is written only by analyze. Canary failures and run/resume integrity or ambiguity failures raise without writing P1-X. error_max_budget_usd is classified as valid_task_failure_v1, so retries and later provider calls continue. Persist claim-disabled P1-X before returning from every stop path, and block further provider calls after spend failure.

🤖 Prompt for AI Agents 
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@research/p1-live-authorization-packet.md` around lines 117 - 120, Update the
v8 stop-path handling so every terminal failure, including canary, run/resume
integrity or ambiguity, and error_max_budget_usd spend failures, persists the
canonical claim-disabled P1-X before returning. Ensure spend failure is not
treated as valid_task_failure_v1: stop retries and block all subsequent provider
calls, while preserving the existing study-invalid-decision.json behavior where
applicable.

- PR #298 merged the provider-free P2-P6 evaluator as `d4b6302...`; candidate
run `31457488674` passed for that exact merge and its downloaded manifest and
tarballs passed paired smoke and attestation verification. Because this
authorization amendment changes the selected source revision, that earlier
candidate is delivery evidence only and cannot be reused for v8. A new
reviewed merge, retained ref, and exact candidate are required before live
`prepare`.
Comment on lines +111 to +127

@coderabbitai coderabbitai Bot Aug 11, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🗄️ Data Integrity & Integration | 🟠 Major | 🏗️ Heavy lift

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
set -euo pipefail

ast-grep outline context-guard-kit/benchmark_runner.py \
  --match '_benchmark_study_v2_invalid_canary_decision' \
  --view expanded

rg -n -C 8 \
  'MERGED_SHA|git ls-remote|manifest|checksums|SRI|provenance|clean install|prepare' \
  --glob '*.py' --glob '*.yml' --glob '*.yaml' .

Repository: ictechgy/context-guard

Length of output: 50379

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '%s\n' '--- benchmark study symbols ---'
rg -n \
  'def (_benchmark_study_v2_invalid_canary_decision|.*prepare.*|prepare)|study-v2-action|candidate|MERGED_SHA|git ls-remote|provenance|SRI|checksum|clean install' \
  context-guard-kit/benchmark_runner.py \
  context-guard-kit \
  scripts \
  tests \
  --glob '*.py' --glob '*.yml' --glob '*.yaml' \
  | head -n 500

printf '%s\n' '--- candidate-related files ---'
git ls-files | rg -i 'benchmark|study|candidate|manifest|provenance|attestation|install'

printf '%s\n' '--- relevant function outlines ---'
ast-grep outline context-guard-kit/benchmark_runner.py | rg -i 'study_v2|prepare|candidate|manifest|provenance|install|canary'

Repository: ictechgy/context-guard

Length of output: 50378

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '%s\n' '--- candidate verification and prepare ---'
sed -n '12040,12655p' context-guard-kit/benchmark_runner.py

printf '%s\n' '--- prepared-manifest revalidation and provider boundary ---'
sed -n '12760,12810p' context-guard-kit/benchmark_runner.py
sed -n '13745,13945p' context-guard-kit/benchmark_runner.py

printf '%s\n' '--- command dispatch ---'
sed -n '14260,14375p' context-guard-kit/benchmark_runner.py

printf '%s\n' '--- candidate fixture and prepare tests ---'
sed -n '930,1100p' scripts/rehearse_measurement_study.py
rg -n -C 6 \
  'prepare_benchmark_study_v2_executable|candidate binding drift|candidate checksum|SRI mismatch|offline npm install|provider' \
  tests/test_benchmark_study_v2.py

Repository: ictechgy/context-guard

Length of output: 50378

🏁 Script executed:

#!/bin/bash
set -euo pipefail

printf '%s\n' '--- candidate manifest production ---'
rg -n -C 10 \
  'commit_sha|candidate_manifest|provenance|attestation|MERGED_SHA|git ls-remote|remote' \
  scripts/build_npm_candidates.py \
  tests/test_npm_candidates.py \
  research/p1-live-authorization-packet.md

printf '%s\n' '--- v2 prepare and candidate-binding tests ---'
rg -n \
  'def test_.*(prepare|candidate|manifest|checksum|integrity|provenance)|prepare_benchmark_study_v2_executable|candidate_manifest_sha256|commit_sha|candidate binding' \
  tests/test_benchmark_study_v2.py \
  tests/test_npm_candidates.py \
  | head -n 300

printf '%s\n' '--- full candidate revalidation tail ---'
sed -n '12700,12805p' context-guard-kit/benchmark_runner.py
sed -n '14335,14380p' context-guard-kit/benchmark_runner.py

Repository: ictechgy/context-guard

Length of output: 40937

Bind commit_sha to the reviewed MERGED_SHA before live prepare.

prepare validates the manifest, checksums, SRI, package contents, provenance fields, and offline clean install. It only checks that commit_sha has valid hex syntax. --study-v2-candidate-hash binds the manifest digest, not the source revision. No runtime check resolves the retained ref with git ls-remote. An old candidate can therefore pass when its manifest digest is supplied. Compare the candidate commit_sha with the approved MERGED_SHA and validate the retained ref before provider execution.

🤖 Prompt for AI Agents 
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@research/p1-live-authorization-packet.md` around lines 111 - 127, Before live
prepare, require the candidate commit_sha to exactly match the approved
MERGED_SHA and validate that the retained ref resolves to that reviewed commit,
rather than relying only on hex syntax or the manifest digest. Update the
candidate validation flow around prepare and --study-v2-candidate-hash,
preserving existing manifest and provenance checks while rejecting mismatched or
unresolved refs before provider execution.


## Frozen study shape

Expand Down Expand Up @@ -140,12 +158,12 @@ estimate or a request to spend that amount. The user may authorize a smaller
amount only by approving a new frozen plan; silently truncating this plan would
make the complete population unreachable and keep P1 failed.

The approved cumulative ceiling is exactly 307 identities. One hundred seventy
identities are already terminally consumed/reserved, leaving 137, which is not
enough for another fresh maximum-size root. No further provider process is
authorized under this packet. No identity beyond the ceiling is authorized, and
the implementation must stop immediately on the frozen integrity, privacy,
ambiguity, or spend conditions rather than trying to consume the ceiling.
The approved cumulative ceiling is exactly 388 identities. One hundred seventy
identities are already terminally consumed/reserved, leaving exactly 218 for
one fresh maximum-size root. No identity beyond the ceiling and no second fresh
root are authorized. The implementation must stop immediately on the frozen
integrity, privacy, ambiguity, or spend conditions rather than trying to
consume the ceiling.

## Frozen Claude executable identity

Expand Down Expand Up @@ -258,17 +276,17 @@ decision.
- [x] GitHub network use for fetch/push/PR/CI, limited to
`ictechgy/context-guard`.
- [x] Claude live study authority: executable above, `sonnet`, cumulative
ceiling exactly 307 consumed/reserved identities, `$0.75` per-process
ceiling exactly 388 consumed/reserved identities, `$0.75` per-process
CLI caps, and explicit exact-CLI internal reuse of the bound existing
first-party login under the prompt/data boundary and stop rules above.
One hundred seventy identities are terminally accounted; only 137 remain,
so this packet authorizes no new root or further provider process.
One hundred seventy identities are terminally accounted; exactly 218
remain for one fresh v8 root under the frozen shape and stop rules.
- [x] npm candidate workflow for the exact approved commit.
- [ ] npm publication of the exact pair to `next`.
- [ ] npm promotion of that exact reviewed pair to `latest`.

The original approvals were recorded on 2026-08-09; auth reuse and the v6
ceiling were recorded on 2026-08-10; the exact v7 ceiling was recorded on
2026-08-11. `next` and `latest` remain blocked. No credential value may be shown
ceiling were recorded on 2026-08-10; the exact v7 and v8 ceilings were recorded
on 2026-08-11. `next` and `latest` remain blocked. No credential value may be shown
to or inspected by Codex; the exact CLI consumes the existing login internally
without exposing values.
28 changes: 20 additions & 8 deletions research/token-savings-roadmap.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,7 +18,7 @@ Current position:
| --- | --- |
| Shipped-code readiness | Narrow P3-style Bash reference route is merged; closed provider-free P2-P6 evaluators are implemented but non-activating. |
| Evidence readiness | P1-X: v7 stopped safely after 79 analytic identities; P1-F is not passed. |
| Release readiness | Earlier attested candidates exist; the P2-P6 evaluator change still requires a new exact candidate. npm publication has not occurred. |
| Release readiness | The P2-P6 evaluator merge and candidate passed; the newly recorded v8 live authorization requires a new exact candidate. npm publication has not occurred. |
| Public claim | Forbidden: the provider-backed v7 decision is P1-X and claim-disabled. |

The roadmap is dependency-gated:
Expand Down Expand Up @@ -240,6 +240,15 @@ is diagnostic-only and cannot replace workflow attestation.
failures remain infrastructure-invalid. The frozen v7 artifact is recognized
read-only as 453,278 primary tokens, but its v7 decision remains P1-X; only a
fresh reviewed candidate and fresh root may exercise the new contract.
- The user approved the exact v8 retry boundary on 2026-08-11: cumulative
ceiling 388, retaining the 170 prior terminal identities and authorizing at
most 218 new identities in one fresh root. The frozen shape remains two
canaries + 108 initials + up to 108 policy-valid retries, `sonnet`, `$0.75`
per process, `$163.50` arithmetic maximum, no optional stopping, no old-root
reuse, and immediate P1-X on integrity, privacy, ambiguity, or spend failure.
PR #298 merge `d4b6302...` and candidate run `31457488674` verified the
provider-free P2-P6 delivery, but the authorization amendment requires a new
reviewed merge and exact candidate before live v8 `prepare`.

P2-P6 active promotion remains stopped until a fresh P1 exit gate passes.

Expand Down Expand Up @@ -370,13 +379,16 @@ claim authority remain false, and exact fallback is unchanged. Delivery may
proceed only to a GitHub PR or an npm candidate under separately recorded
authority; npm `next` and `latest` remain untouched.

1. Finish package integrity, release smoke, full prepublish, review, and the
GitHub PR for the provider-free P2-P6 evaluator.
2. Build and attest a new exact npm candidate only. Keep npm `next`, npm
`latest`, artifact deletion, and all P2-P6 provider work blocked.
3. Keep active promotion stopped until a fresh P1-F exit exists. The current
P1-X evidence and local evaluator outputs unlock nothing.
4. After P1-F, obtain a new phase-specific call/spend/privacy authorization and
1. Review and merge the exact v8 authorization amendment, retain its exact
source ref, then build and attest a new npm candidate only.
2. Run provider-free prepare and verify the exact CLI, existing-login safe
projection, candidate overlay, fixtures, settings, runtime, and fresh private
root before either live canary.
3. Execute the fixed v8 canaries and analytic schedule under the exact 388
cumulative ceiling. Any stop condition closes the root and yields P1-X.
4. Keep active promotion stopped until a fresh P1-F exit exists. The current
v7 P1-X and local evaluator outputs unlock nothing.
5. After P1-F, obtain a new phase-specific call/spend/privacy authorization and
execute P2 through P6 sequentially; a failed gate stops later phases and
preserves the exact prior route.

Expand Down