Cycle: 19th audit of this cadence
Auditor: SCOUT (TITAN research agent)
Baseline: F:/TITAN/plans/advisors/CLAUDE-CODE-ARCHITECTURE-DEEP-DIVE-2026-04-22.md
Prior audit: F:/TITAN/plans/advisors/claude-code-audit-2026-04-27-0419.md (cycle 18, v2.1.119, 0 regressions, T064-T066 filed)
CC version at prior audit: v2.1.119
CC version this cycle: v2.1.119 (confirmed via GitHub releases page direct fetch — no v2.1.120 listed; rollback stands)
Local TITAN install: v2.1.49 (70-version gap; T030 open, ceiling pinned at v2.1.119 per T052)
Next unclaimed T-numbers: T067, T068, T069
Word count: ~2,350
---
Finding: Confirmed (primary source: github.com/anthropics/claude-code/releases, fetched 2026-04-27).
The GitHub releases page — the authoritative primary source for CC releases — lists v2.1.119 (April 23, 2026) as the most recent release. v2.1.120 does not appear on the GitHub releases page. This confirms the T052 ceiling assessment: v2.1.120 was rolled back and has not been re-issued. The WebSearch snippet from cycle 18 claiming "v2.1.120 published 2 hours ago" is definitively stale search index data.
T064 implication. The T064 tripwire condition — "T052 ceiling re-evaluates when gist.github.com/yurukusa/a866b4cd2976486156a00c190c39cef6 reports all 8 regressions closed OR Anthropic publishes explicit re-release note" — remains active. The GitHub releases page is now added as a standing primary source alongside releasebot.io for version checks.
No new releases shipped in the 3.5-hour window between cycle 18 and cycle 19. The v2.1.118-v2.1.119 architectural changes captured below are first fully analyzed this cycle from multiple primary sources (GitHub changelog, releasebot.io, code.claude.com/docs/en/changelog).
Sources: github.com/anthropics/claude-code/releases (fetched 2026-04-27; v2.1.119 confirmed latest; v2.1.120 absent); releasebot.io/updates/anthropic/claude-code (fetched 2026-04-27; consistent); code.claude.com/docs/en/changelog (fetched 2026-04-27).
---
Finding: Confirmed (code.claude.com/docs/en/changelog, v2.1.118 entry, fetched 2026-04-27). Not present in baseline memo or cycles 1-18.
v2.1.118 introduced the ability for hooks to invoke MCP tools directly via type: "mcp_tool" in the hook output. Previously, hooks were limited to: blocking tool calls (permissionDecision: block), modifying tool input parameters (updatedInput), injecting context (additionalContext), and forcing loop termination (exit code 2). The new type: "mcp_tool" hook output allows a hook firing on PreToolUse or PostToolUse to trigger an MCP tool call as a side effect of the hook evaluation.
Architectural significance. This turns hooks from passive interceptors into active orchestrators. A PreToolUse hook that detects a dangerous file path can now not only block the tool call but simultaneously invoke an MCP notification tool to alert the user. A PostToolUse hook that observes a test failure can invoke an MCP logging tool to persist the failure context. This is a second-order orchestration capability that the baseline memo did not anticipate.
TITAN implication. TITAN's hook architecture (currently: PreToolUse blocks, PostToolUse logs) can be extended to invoke MCP notification tools as part of hook responses. The most immediate application: TITAN's audit routine could use a PostToolUse hook to push an MCP notification when an audit writes a new memo file, rather than requiring a separate notification step.
SI implication. None — SI does not use hooks or MCP.
Sources: code.claude.com/docs/en/changelog (v2.1.118 entry, "Hooks can now invoke MCP tools directly via type: 'mcp_tool'", fetched 2026-04-27).
---
Finding: Confirmed (code.claude.com/docs/en/changelog, v2.1.117 entry, fetched 2026-04-27). First full analysis this cycle.
v2.1.117 enabled forked subagents on external builds via the CLAUDE_CODE_FORK_SUBAGENT=1 environment variable. The baseline memo (section 2.4) described three subagent isolation modes — in-process, worktree, and remote — but noted that worktree and remote modes were behind feature() build-time gates unavailable in external builds. The v2.1.117 flag opens forked subagent mode to external (non-Anthropic) installations.
Additionally: agent frontmatter mcpServers is now loaded for main-thread agent sessions via --agent, meaning a named agent definition can carry its own MCP server configuration independent of the user's global settings.
Architectural significance. Forked subagents use filesystem isolation (Git worktrees) and return only summaries to the parent context. The baseline described this as an AWS Lambda analogy — each agent gets a fresh context window, constrained tools, and leaves no verbosity in the parent. Making this available externally means TITAN can spawn true filesystem-isolated subagents rather than the current in-process pattern. The agent frontmatter mcpServers binding means each named agent can have its own specialized MCP toolset, enabling purpose-built agents with scoped capabilities.
TITAN implication. T030 (upgrade from v2.1.49 to v2.1.119) unlocks forked subagent capability. TITAN's six named agents (SCOUT, VAULT, FORGE, GUIDE, ORACLE, DARWIN) could be promoted to worktree-isolated agents via CLAUDE_CODE_FORK_SUBAGENT=1, giving each a clean filesystem context and preventing cross-contamination of session state. This is a significant capability upgrade relative to the current in-process model.
SI implication. None — SI does not use CC subagents.
Sources: code.claude.com/docs/en/changelog (v2.1.117 entry, "Forked Subagents: Now enabled on external builds by setting CLAUDE_CODE_FORK_SUBAGENT=1"; "Agent MCP: Agent frontmatter mcpServers now loaded for main-thread agent sessions via --agent", fetched 2026-04-27).
---
Finding: Confirmed (devtoolpicks.com/blog/anthropic-claude-code-quality-fix-postmortem-2026, fetched 2026-04-27; cross-referenced with HN item 47878905, fetched 2026-04-27).
Anthropic's postmortem confirmed three distinct product-layer configuration changes that degraded CC quality between March 4 and April 20, 2026. Model weights were unchanged throughout. All three were configuration-only regressions:
Regression 1 — Reasoning Effort Reduction (March 4, reverted April 7): Default reasoning effort shifted from high to medium to improve UI responsiveness. Result: shallower analysis, simplistic fixes on complex tasks. This directly contradicts the baseline memo's framing of CC as a system that "reasons deeply before acting." v2.1.118 explicitly reversed this, restoring high as the default effort for Pro/Max on Opus 4.6/Sonnet 4.6.
Regression 2 — Caching Bug (March 26, fixed April 10): Session management defect caused Claude to clear cached thinking every turn instead of once after idle periods. Users experienced "forgetful and repetitive" behavior on resumed sessions. This is the compaction layer failing: the context was being degraded not by the five-layer graduated pipeline but by a runaway cache-clearing loop.
Regression 3 — Verbosity System Prompt (April 16, reverted April 20): A system prompt instruction was added: "keep text between tool calls to 25 words or less." While internal testing appeared safe, broader ablations found a 3% coding performance drop. This is a direct example of an anti-pattern for SI: system prompt verbosity constraints that produce measurable quality degradation.
SI architectural signal. Regression 3 is the most directly relevant to Silent Infinity. The verbosity constraint ("25 words or less between tool calls") degraded CC performance by 3%. In a wellness product, an analogous constraint (e.g., "responses must be under 100 words") could similarly degrade the contemplative quality of responses. The CC postmortem confirms that brevity constraints require proper ablation testing against quality metrics before shipping — not just UI-side testing. This reinforces the anti-pattern documented in the baseline (AP-3: committed/confident declarative tone is a coding-tool convention, not a wellness-product virtue).
Sources: devtoolpicks.com/blog/anthropic-claude-code-quality-fix-postmortem-2026 (fetched 2026-04-27); news.ycombinator.com/item?id=47878905 (fetched 2026-04-27).
---
Finding: Confirmed (code.claude.com/docs/en/changelog, v2.1.117 entry, fetched 2026-04-27).
v2.1.117 corrected Opus 4.7's context window from the assumed 200K tokens to its actual 1M token native window. The /context display command was also fixed to compute percentages against the correct 1M limit. Prior audit cycles used 200K as the Opus 4.7 context assumption. This is superseded.
TITAN implication. TITAN's ULTRAPLAN pattern (baseline memo section 1.4 — "runs Opus 4.6 with up to 30 minutes of dedicated think time") would benefit from Opus 4.7's 1M context if deep-planning sessions were to exceed the 200K boundary. At current usage levels, this is unlikely to matter. The implication to note is that context window management heuristics based on "Opus = 200K" are now stale and should use "Opus 4.7 = 1M" when determining compaction trigger thresholds.
SI implication. SI uses Bedrock; Opus 4.7 availability on Bedrock requires separate confirmation (not covered this cycle).
Sources: code.claude.com/docs/en/changelog (v2.1.117 entry, "Context Window: Opus 4.7 support with 1M context window (vs. 200K assumed previously)", fetched 2026-04-27).
---
Finding: Confirmed (glob scan C:/Users/Harnoor/.claude/skills/, 2026-04-27).
The skills directory contains 13 skills across TITAN's named workflows: sense/, evolve/, pulse/, monologue/, reflect/, newsletter/, teach/, learn/, titan/, briefing/, feed/, dream/, plus sense/token-tracker.md. No new skills were added since cycle 18. The ~/.claude/hooks/ directory is absent (confirmed via glob error — directory does not exist). No new MCP server configurations detected.
Observation. TITAN has 13 skills but no hooks directory. The v2.1.118 capability (hooks invoking MCP tools) cannot be leveraged without creating a hooks configuration. This is a low-hanging capability gap given the existing skill infrastructure. See Recommendation AT (T067).
---
Status: No new SI production deployments detected since cycle 18. Gap table carries forward from cycle 18 with one change: Pattern 8 classification refined.
| # | Pattern | CC Baseline | SI Status | Gap |
|---|---------|------------|-----------|-----|
| 1 | Memory layering (hot/warm/cold) | MEMORY.md file-tiered | ALIGNED — DDB 4-tier memory.py live | None |
| 2 | System prompt composition (conditional stack) | 6-layer conditional | ALIGNED — versioned + variant + user context injection | None |
| 3 | Structured tool use (schema-validated) | 50 tools, JSON Schema | GAP — capabilities in prose, not formal tool schemas | T025 open |
| 4 | Sub-agent orchestration | Named agents, forked subagents now external (v2.1.117) | PARTIAL — Chat Sentinel exists; no parallel workers | Partial |
| 5 | Verification-before-claim | Harness validates tool results | ALIGNED — system prompt discipline instruction live | None |
| 6 | Plan mode / reflective pause | Shift+Tab read-only posture | PARTIAL — contemplative persona exists; no explicit mode | Partial |
| 7 | Correction-as-memory | Live feedback → persistent rules | ALIGNED — extract_correction() → memory.put_correction() live | None |
| 8 | Skill auto-invocation (domain injection) | Semantic match, lazy-load | PARTIAL — skills_loader.py behind SKILLS_ENABLED=1; manifest unconfirmed | T046 + T065 open |
| 9 | Session transcript rehydration on reconnect | JSONL + /recap + /fork + 67% faster resume (v2.1.116) | PARTIAL — recap wired; no fork endpoint; returning-user UX absent | T060 open |
| 10 | Interruptible streaming / barge-in | ESC mid-stream + partial transcript | PARTIAL — SSE abort at Lambda; no client interrupt UX | Partial |
| 11 | Memory compaction (graduated pipeline) | 5-layer cheapest-first; caching bug fixed (v2.1.116) | ALIGNED — 2-layer compaction in conversation_store.py | None |
| 12 | Permission / guardrail model (deny-first) | disableSkillShellExecution; 8 security layers | ALIGNED — guardrails.py + Haiku classifier | None |
| 13 | Pre-session briefing (context injection) | SessionStart hook + managed-settings.d/ fragments | ALIGNED — memory block injected as late user message (T014 closed) | None |
| 14 | Parallel tool calls | StreamingToolExecutor + parallel MCP reconfiguration (v2.1.119) | GAP — single-threaded Lambda; asyncio.gather() partially mitigates | T051 open |
Regressions this cycle: 0. No SI deployments detected. No new gaps opened.
Pattern 14 update. v2.1.119 extended parallel execution: subagent MCP server reconfiguration now connects servers in parallel instead of serially. This is an incremental parallel execution expansion in CC, deepening the gap relative to SI's single-threaded Lambda model. T051 remains the open task.
Pattern 4 update. v2.1.117's external forked subagent support (CLAUDE_CODE_FORK_SUBAGENT=1) represents a meaningful CC capability expansion in this pattern. SI's PARTIAL status (Chat Sentinel exists; no parallel workers) is unchanged but the gap relative to CC's Pattern 4 capability has grown. See Recommendation AU (T068).
---
Next unclaimed T-numbers: T067, T068, T069.
---
Problem. TITAN has 13 skills and zero hooks (~/.claude/hooks/ does not exist). v2.1.118 added the ability for hooks to invoke MCP tools directly (type: "mcp_tool"). This means TITAN could wire a PostToolUse or Stop hook to push an MCP notification when an audit routine completes — replacing the current pattern of requiring a separate explicit notification step after audit memo creation.
The practical gap: audit memos are written to F:/TITAN/plans/advisors/ by SCOUT via Write tool, but the audit-completion notification (email via create_draft MCP) currently requires an explicit agent step. If the Write tool fires a PostToolUse hook that checks whether the written file is an audit memo (path pattern: F:/TITAN/plans/advisors/claude-code-audit-*.md), the hook can invoke the create_draft MCP tool directly, eliminating the notification as an explicit task step.
This is a direct application of the v2.1.118 hook-invokes-MCP-tool capability to TITAN's own automation.
Fix — under 4 hours:
1. Create ~/.claude/hooks/post_tool_use.sh (or .py) that reads the hook input JSON from stdin.
2. If hook_event_name == "PostToolUse" and tool_name == "Write" and the written file path matches /advisors/claude-code-audit-.md:
- Extract the memo path, n_recs, m_regressions from the audit memo content (simple grep for "n_recs" and "m_regressions" lines from the last cadence log entry, or from the memo title block).
- Output {"type": "mcp_tool", "server": "gmail", "tool": "create_draft", "input": {...}} with the digest email fields populated.
3. Register the hook in ~/.claude/settings.json under hooks.PostToolUse.
4. Test: write a dummy claude-code-audit-test.md file, confirm hook fires and draft appears in Gmail.
Why this is Recommendation 1. It closes a genuine automation gap (audit notification requires explicit agent action rather than being triggered by the memo write), uses a new CC capability (v2.1.118 hook-to-MCP) that TITAN is currently not using, and is under 4 hours. The hooks directory being absent is the most obvious gap between TITAN's stated capability and its installed configuration.
Blast radius: New ~/.claude/hooks/ directory and post_tool_use.sh. Settings.json hook registration. Zero code changes to SI. Zero changes to existing skills.
Effort: 3-4 hours (LOW)
Priority: HIGH — closes automation gap; uses new CC capability; under 1 day of work
Dependencies: create_draft MCP server must be accessible from hook context (verify MCP auth token is available in hook execution environment before wiring)
File as T067.
Sources: code.claude.com/docs/en/changelog (v2.1.118, "Hooks can now invoke MCP tools directly via type: 'mcp_tool'", fetched 2026-04-27); glob scan C:/Users/Harnoor/.claude/hooks/ (absent, confirmed 2026-04-27); glob scan C:/Users/Harnoor/.claude/skills/ (13 skills confirmed, fetched 2026-04-27).
---
Problem. T030 tracks upgrading TITAN's local CC from v2.1.49 to v2.1.119. The T030 entry was written when forked subagent support was unavailable on external builds (baseline memo section 2.4 noted it as behind a build-time gate). v2.1.117 opened forked subagents externally via CLAUDE_CODE_FORK_SUBAGENT=1. This is a meaningful capability upgrade that T030 should explicitly capture.
TITAN's six named agents (SCOUT, VAULT, FORGE, GUIDE, ORACLE, DARWIN) currently run as in-process subagents — they share the parent session context and can contaminate each other's state. Forked subagents use Git worktree isolation: each agent gets a clean filesystem context, cannot see the parent's tool results, and returns only a summary. This is the "Lambda analogy" from the baseline memo applied to TITAN's own agent architecture.
The gap is not about shipping something new — it is about ensuring T030's upgrade plan captures the post-upgrade configuration step to enable CLAUDE_CODE_FORK_SUBAGENT=1 and optionally promotes the six named agents to worktree-isolated runs.
Fix — 30 minutes (annotation only):
1. Annotate T030 with: "Post-upgrade step: evaluate CLAUDE_CODE_FORK_SUBAGENT=1 for TITAN named agents. If enabled, each of the 6 named agents (SCOUT, VAULT, FORGE, GUIDE, ORACLE, DARWIN) runs in a worktree-isolated context with a clean filesystem boundary. Benefit: prevents cross-agent context contamination. Risk: worktree isolation requires each agent's session to start clean — agent frontmatter mcpServers must be configured per agent. Reference: v2.1.117 changelog. Prerequisite: confirm Git is available in the TITAN working directory (required for worktree mode)."
2. Note the dependency: forked subagents require Git in the working directory. Confirm F:/TITAN is a git repo or that the working directory used for agent spawning is.
Why this is Recommendation 2. T030 will execute at some point (it has been open since the registry was created). Without this annotation, the upgrade would proceed without capturing a significant new capability that directly applies to TITAN's architecture. The annotation itself is 30 minutes; the capability unlocked is substantial. This is a leverage multiplier on an already-planned task.
Blast radius: Task registry annotation on T030. Zero code changes. Zero SI impact.
Effort: 30 minutes (TRIVIAL — annotation)
Priority: MEDIUM — needed before T030 executes; T030 is currently open with no confirmed execution date
Dependencies: Should complete before T030 executes. Complements T064 (version tripwire) and T066 (DISABLE_UPDATES).
File as T068.
Sources: code.claude.com/docs/en/changelog (v2.1.117, "Forked Subagents: Now enabled on external builds by setting CLAUDE_CODE_FORK_SUBAGENT=1"; "Agent MCP: Agent frontmatter mcpServers now loaded for main-thread agent sessions via --agent", fetched 2026-04-27); baseline memo section 2.4 (subagent isolation modes; worktree mode previously gated); TASK-REGISTRY-2026-04-21.md T030 entry (read 2026-04-27).
---
Problem. The CC quality postmortem (March-April 2026) confirmed that a verbosity system prompt instruction ("keep text between tool calls to 25 words or less") caused a 3% coding performance regression that only appeared in broad ablations — not in the initial internal testing that approved the change for ship. The instruction was reverted 4 days after ship.
Silent Infinity's Feature Readiness Standard governs what must be true before a feature ships to users. It currently mandates 582+ green tests and a Feature Readiness Checklist. It does not explicitly require response-quality ablation testing for system prompt changes. This is the same gap that allowed the CC verbosity regression to reach production.
SI's risk profile is higher than CC's here: CC operates on coding tasks where a 3% performance drop in code quality is measurable and reversible. SI operates on vulnerable users in wellness conversations where a response quality drop — shorter, terser, less contemplative — could degrade the user's felt safety without any measurable error signal in the test suite. The postmortem's lesson is: system prompt changes that affect response length, tone, or reasoning depth require behavioral quality metrics, not just functional correctness tests.
Fix — under 1 day:
1. Add a new section to SI's Feature Readiness Standard: "System Prompt Quality Gate." Contents: before shipping any system prompt change that affects (a) response length targets, (b) verbosity instructions, (c) reasoning effort hints, or (d) tone directives, a blind evaluation must be run: 10 representative conversation turns, two variants (current vs. proposed), blind rating by Harnoor on contemplative quality (1-5). The proposed variant must score >= current variant on average before shipping. This is the CC-postmortem-informed ablation protocol.
2. The blind evaluation is the minimum bar. It requires: 2 hours to generate representative turns, 30 minutes to rate, 30 minutes to document. Total overhead per system prompt change: ~3 hours. Cost: eliminates the class of regressions where "looks fine in unit tests, degrades in practice."
3. Add a "postmortem-informed" annotation to the checklist item referencing the CC quality incident as the motivation.
Why this is Recommendation 3. The CC postmortem is the most directly actionable new intelligence this cycle for SI. The three regression types — reasoning effort reduction, caching bugs, verbosity constraints — map cleanly to risks that SI faces on every system prompt iteration. Documenting the ablation protocol in the Feature Readiness Standard costs under 1 day and prevents an entire class of silent quality regressions from reaching SI's users. This is also an anti-pattern documentation task: the postmortem gives TITAN a concrete "what not to ship" reference for future SI development cycles.
Blast radius: Feature Readiness Standard document update only. Zero production code changes. Zero infrastructure changes.
Effort: 3-4 hours (LOW — documentation only)
Priority: HIGH — prevents quality regression class; grounded in confirmed CC postmortem; applies to every future SI system prompt change
Dependencies: Access to SI's Feature Readiness Standard document path (not confirmed this cycle — TITAN to locate before executing).
File as T069.
Sources: devtoolpicks.com/blog/anthropic-claude-code-quality-fix-postmortem-2026 (Regression 3 — verbosity system prompt, "25 words or less," 3% performance drop, reverted April 20, fetched 2026-04-27); news.ycombinator.com/item?id=47878905 (community reaction — silent degradation without user notification, fetched 2026-04-27); baseline memo Anti-Pattern 3 (committed/confident tone is a coding-tool convention, not a wellness-product virtue); F:/TITAN/plans/advisors/CLAUDE-CODE-ARCHITECTURE-DEEP-DIVE-2026-04-22.md (baseline; Feature Readiness Standard referenced in section on "How to port to Silent Infinity").
---
AP-11 — Shipping System Prompt Changes Without Behavioral Ablation Testing.
The CC quality postmortem confirms: three configuration-only changes degraded model quality in ways that internal testing did not catch. The failure mode is: functional tests (does the tool call succeed? does the response arrive?) do not measure behavioral quality (is the response as good as before?). For SI, this anti-pattern is more dangerous than for CC because SI's quality signal is contemplative depth — a property that existing test suites do not measure. AP-11 is the motivation for Recommendation AV (T069).
AP-10 reinforcement (version monitoring via search snippets): The T064 cycle-18 incident is now resolved — the GitHub releases page confirms v2.1.120 has not been re-issued. The primary-source fetch discipline established in T064 prevented a false-positive upgrade decision.
AP-1 through AP-9 (carried from prior cycles; no changes): See F:/TITAN/plans/advisors/claude-code-audit-2026-04-27-0419.md Section 4 for the full list.
---
Resolution: v2.1.120 status confirmed resolved. The cycle-18 contradiction (WebSearch snippet claimed v2.1.120 published "2 hours ago" vs. releasebot.io showing v2.1.119 as latest) is now resolved. GitHub releases page — the primary source — confirms v2.1.120 is not listed. Rollback stands. T052 ceiling is correct. T064 remains open to formalize the primary-source fetch protocol for future cycles.
Uncertainty 1 — skills_loader.py manifest content (T046, T065): Unchanged from cycle 18. Pattern 8 PARTIAL. T065 makes this the explicit next-action.
Uncertainty 2 — SI Feature Readiness Standard document path: T069 requires updating this document. Its path is not confirmed from TITAN's current knowledge. TITAN should locate the document before T069 executes.
Uncertainty 3 — MCP auth token availability in hook execution context: T067 (audit-completion hook) requires that the create_draft MCP tool is accessible from within a hook's execution environment. Hook scripts run as shell processes; MCP auth tokens may or may not be available in that context depending on how TITAN's MCP servers are configured. This must be verified before T067 is executed. If auth tokens are not available in hook context, T067's approach shifts to a simpler curl-to-smtp or python send_mail alternative.
Uncertainty 4 — Opus 4.7 availability on Bedrock: SI uses AWS Bedrock for model inference. v2.1.117's 1M context window discovery applies to Opus 4.7, but whether Opus 4.7 is available on Bedrock with the full 1M context is unconfirmed. Requires a separate Bedrock API check before any SI compaction thresholds are recalibrated.
---
| Item | Count |
|------|-------|
| CC versions reviewed (cumulative since baseline) | 22 (v2.1.89 through v2.1.119; v2.1.120 rolled back confirmed) |
| New CC architectural signals this cycle | 3 (hook→MCP invocation v2.1.118; forked subagents external v2.1.117; Opus 4.7 1M context v2.1.117) |
| Quality postmortem items analyzed | 3 (reasoning effort reduction, caching bug, verbosity constraint — all from March-April 2026) |
| SI regressions detected | 0 |
| SI positive developments | 0 new deployments detected |
| Persistent SI pattern gaps | Pattern 3 (T025), Pattern 4 partial, Pattern 6 partial, Pattern 8 partial (T046+T065), Pattern 9 partial (T060), Pattern 10 partial, Pattern 14 (T051) |
| New recommendations filed | 3 (T067: audit-completion hook; T068: forked subagent annotation; T069: verbosity ablation protocol) |
| Anti-patterns documented (cumulative) | 11 (AP-1 through AP-11; AP-11 added this cycle from quality postmortem) |
| Open T-numbers with direct SI impact | T025, T028, T037, T038, T040, T041, T046, T047, T048, T051, T053, T056, T060, T061, T062, T065, T069 |
| Open T-numbers with TITAN-only impact | T026, T029, T030, T031, T032, T033, T034, T035, T036, T039, T042, T043, T044, T045, T049, T050, T052, T054, T055, T057, T058, T059, T063, T064, T066, T067, T068 |
---
1. F:/TITAN/plans/advisors/CLAUDE-CODE-ARCHITECTURE-DEEP-DIVE-2026-04-22.md — baseline memo (SCOUT, 2026-04-22; read 2026-04-27)
2. F:/TITAN/plans/advisors/claude-code-audit-2026-04-27-0419.md — cycle 18 prior audit (read 2026-04-27; last T-number T066)
3. F:/TITAN/plans/task-registry/TASK-REGISTRY-2026-04-21.md — task registry (read 2026-04-27; T066 confirmed last T-number)
4. F:/TITAN/plans/audit-cadence.log — audit history (read 2026-04-27; last completed entry 2026-04-26T22:19:00Z; dispatched entry 2026-04-27T07:52)
5. github.com/anthropics/claude-code/releases — authoritative CC release page (fetched 2026-04-27; v2.1.119 confirmed latest; v2.1.120 absent — rollback confirmed)
6. releasebot.io/updates/anthropic/claude-code — release aggregator April 2026 (fetched 2026-04-27; consistent with GitHub; v2.1.118 DISABLE_UPDATES confirmed; v2.1.119 changelog confirmed)
7. code.claude.com/docs/en/changelog — official CC changelog (fetched 2026-04-27; v2.1.116-v2.1.119 architectural details; hook→MCP v2.1.118; forked subagents v2.1.117; Opus 4.7 1M context v2.1.117; effort default high v2.1.118)
8. devtoolpicks.com/blog/anthropic-claude-code-quality-fix-postmortem-2026 — quality incident postmortem (fetched 2026-04-27; three regressions confirmed; verbosity constraint AP-11 source)
9. news.ycombinator.com/item?id=47878905 — HN: "An update on recent Claude Code quality reports" (fetched 2026-04-27; community reaction; caching bug details confirmed)
10. Glob scan C:/Users/Harnoor/.claude/skills/ (fetched 2026-04-27; 13 skills confirmed; hooks directory absent)
11. WebSearch results (retrieved 2026-04-27; secondary source; used for link discovery only — primary source fetches took precedence for factual claims)