alex/g3
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
1b26de6cd2170be8495593efa888a1776311f452
g3/analysis/memory.md
Dhanji R. Prasanna 1b26de6cd2 Fix context window progress bar showing wrong token counts 
Calibrate used_tokens from API prompt_tokens (ground truth) to fix
progress bar drift in interactive mode. Three issues fixed:

1. update_usage_from_response() only updated cumulative_tokens, never
   calibrated used_tokens. Now snaps used_tokens to prompt_tokens when
   available (falls back to heuristic when prompt_tokens is 0).

2. Moved calibration call inline during streaming (when usage chunk
   arrives) instead of after the loop. Text-only responses — the most
   common case in interactive mode — take an early return path that
   bypassed the post-loop usage update entirely.

3. Removed mock Usage with hardcoded prompt_tokens=100 from
   execute_single_task() which corrupted calibration.
2026-03-18 15:31:20 +11:00

25 KiB
Raw Blame History

Workspace Memory

Updated: 2026-03-18T03:59:01Z | Size: 25.2k chars

Remember Tool Wiring

  • crates/g3-core/src/tools/memory.rs [0..5686]
    • get_memory_path() [486] - resolves analysis/memory.md
    • execute_remember() [1066] - tool handler
    • merge_memory() [2324] - merges new notes into existing
  • crates/g3-core/src/tool_definitions.rs [956..] - remember tool in create_core_tools()
  • crates/g3-core/src/tool_dispatch.rs [670] - dispatch case
  • crates/g3-core/src/prompts.rs [4200..6500] - Workspace Memory prompt section
  • crates/g3-cli/src/project_files.rs - read_workspace_memory() loads analysis/memory.md

Context Window & Compaction

  • crates/g3-core/src/context_window.rs [0..43282]
    • ThinResult [765] - scope, before/after %, chars_saved
    • ContextWindow [2220] - token tracking, message history
    • add_message_with_tokens() [3171] - preserves messages with tool_calls even if content empty
    • estimate_message_tokens() [7695] - sums content + tool_calls[].input tokens (chars/3 * 1.1 + 20 overhead)
    • should_compact() [8954] - threshold check (80%)
    • reset_with_summary() [10685] - compact history to summary
    • reset_with_summary_and_stub() [11120] - ACD integration
    • extract_preserved_messages() [13199] - strips tool_calls from last assistant to prevent orphaned tool_use
    • thin_context() [15038] - replace large results with file refs
  • crates/g3-core/src/compaction.rs [0..11404]
    • CompactionResult, CompactionConfig - result/config structs
    • perform_compaction() - unified for force_compact() and auto-compaction
    • calculate_capped_summary_tokens(), should_disable_thinking()
    • build_summary_messages(), apply_summary_fallback_sequence()
    • ACD integration [195..240] - creates fragment+stub during compaction
  • crates/g3-core/src/lib.rs
    • force_compact() [47902]
    • stream_completion_with_tools() [85389] - main agent loop

Session Storage & Continuation

  • crates/g3-core/src/session_continuation.rs [0..22907]
    • SessionContinuation [1024]
    • save_continuation() [5581]
    • load_continuation() [6428]
  • crates/g3-core/src/paths.rs [0..5498]
    • get_session_logs_dir() [2434]
    • get_thinned_dir() [3060]
    • get_fragments_dir() [3295] - .g3/sessions/<id>/fragments/
    • get_session_file() [3517]
  • crates/g3-core/src/session.rs - session logging utilities

Tool System

  • crates/g3-core/src/tool_definitions.rs [0..15391]
    • ToolConfig [381]
    • create_tool_definitions() [742]
    • create_core_tools() [956]
  • crates/g3-core/src/tool_dispatch.rs [0..3983] - dispatch_tool() [670] routing

CLI Module Structure

  • crates/g3-cli/src/lib.rs [0..9309] - run() [1242], mode dispatch, config loading
  • crates/g3-cli/src/cli_args.rs [0..6043] - Cli [1374] struct (clap)
  • crates/g3-cli/src/autonomous.rs [0..25630] - run_autonomous() [638], coach-player loop
  • crates/g3-cli/src/agent_mode.rs [0..13558] - run_agent_mode() [1000]
  • crates/g3-cli/src/accumulative.rs [0..12006] - run_accumulative_mode() [796]
  • crates/g3-cli/src/interactive.rs [0..19222] - run_interactive() [3809], REPL
  • crates/g3-cli/src/task_execution.rs [0..5520] - execute_task_with_retry() [1069]
  • crates/g3-cli/src/commands.rs [0..20115] - /help, /compact, /thinnify, /fragments, /rehydrate
  • crates/g3-cli/src/utils.rs [0..6154] - display_context_progress(), setup_workspace_directory(), load_config_with_cli_overrides()
  • crates/g3-cli/src/display.rs [0..12573] - format_workspace_path() [286], LoadedContent, print_loaded_status()

Auto-Memory System

  • crates/g3-core/src/lib.rs
    • tool_calls_this_turn [5272] - tracks tools per turn
    • set_auto_memory() [64643] - enable/disable
    • send_auto_memory_reminder() [72800] - MEMORY CHECKPOINT prompt
    • execute_tool_in_dir() [132582] - records tool calls
  • crates/g3-core/src/prompts.rs [3800..4500] - Memory Format in system prompt
  • crates/g3-cli/src/lib.rs - --auto-memory CLI flag

Streaming Markdown Formatter

  • crates/g3-cli/src/streaming_markdown.rs [0..37669]
    • process_in_code_block() [17159] - detects closing fence
    • format_header() [21339] - headers with inline formatting
    • emit_code_block() [27134] - joins buffer, highlights code
    • flush_incomplete() [28434] - handles unclosed blocks at stream end
  • crates/g3-cli/tests/streaming_markdown_test.rs - header formatting tests
  • Gotcha: closing ``` without trailing newline must be detected in flush_incomplete()

Retry Infrastructure

  • crates/g3-core/src/retry.rs [0..11865] - execute_with_retry(), retry_operation(), RetryConfig, RetryResult
  • crates/g3-cli/src/task_execution.rs - execute_task_with_retry()

UI Abstraction Layer

  • crates/g3-core/src/ui_writer.rs [0..8007] - UiWriter trait [211], NullUiWriter [6538], print_thin_result() [1136]
  • crates/g3-cli/src/ui_writer_impl.rs [0..14000] - ConsoleUiWriter, print_tool_compact()
  • crates/g3-cli/src/simple_output.rs [0..1200] - SimpleOutput helper

Feedback Extraction

  • crates/g3-core/src/feedback_extraction.rs [0..22455] - extract_coach_feedback(), try_extract_from_session_log(), try_extract_from_native_tool_call()
  • crates/g3-cli/src/coach_feedback.rs [0..4025] - extract_from_logs() for coach-player loop

Streaming Utilities & State

  • crates/g3-core/src/streaming.rs [0..27241]
    • MAX_ITERATIONS [419] - constant (400)
    • StreamingState [499] - cross-iteration: full_response, first_token_time, iteration_count
    • ToolOutputFormat [1606] - enum: SelfHandled, Compact(String), Regular
    • format_tool_result_summary() [1743], is_compact_tool() [2635], format_compact_tool_summary() [3179]
    • IterationState [5061] - per-iteration: parser, current_response, tool_executed
    • log_stream_error() [8017], truncate_for_display() [10887], truncate_line() [11247]
    • is_connection_error() [21620]

Background Process Management

  • crates/g3-core/src/background_process.rs [0..9048]
    • BackgroundProcessManager [1466], start() [2601], list() [5527], get() [5731], is_running() [5934], remove() [6462]
  • No stop() method — use shell kill <pid>

Unified Diff Application

  • crates/g3-core/src/utils.rs [5000..15000] - apply_unified_diff_to_string(), parse_unified_diff_hunks()
  • Handles multi-hunk diffs, CRLF normalization, range constraints

Error Classification

  • crates/g3-core/src/error_handling.rs [0..19454]
    • ErrorType [5206], RecoverableError [5465] (enum), classify_error() [5972]
  • Priority: rate limit > network > server > busy > timeout > token limit > context length
  • Gotcha: "Connection timeout" → NetworkError (not Timeout) due to "connection" keyword priority

CLI Metrics

  • crates/g3-cli/src/metrics.rs [0..5416] - TurnMetrics, format_elapsed_time(), generate_turn_histogram()

ACD (Aggressive Context Dehydration)

Saves conversation fragments to disk, replaces with stubs.

  • crates/g3-core/src/acd.rs [0..22830]
    • Fragment - new(), save(), load(), generate_stub(), list_fragments(), get_latest_fragment_id()
  • crates/g3-core/src/tools/acd.rs [0..8500] - execute_rehydrate() tool
  • crates/g3-cli/src/lib.rs - --acd flag; /fragments, /rehydrate commands
  • Fragment JSON: fragment_id, created_at, messages, message_count, user_message_count, assistant_message_count, tool_call_summary, estimated_tokens, topics, preceding_fragment_id

UTF-8 Safe String Slicing

Rust &s[..n] panics on multi-byte chars (emoji, CJK) if sliced mid-character. Pattern: s.char_indices().nth(n).map(|(i,_)| i).unwrap_or(s.len()) Danger zones: Display truncation, ACD stubs, user input, non-ASCII text.

Studio - Multi-Agent Workspace Manager

  • crates/studio/src/main.rs [0..12500] - cmd_run(), cmd_status(), cmd_accept(), cmd_discard(), extract_session_summary()
  • crates/studio/src/session.rs - Session, SessionStatus
  • crates/studio/src/git.rs - GitWorktree for isolated agent sessions
  • Session log: <worktree>/.g3/sessions/<session_id>/session.json
  • Fields: context_window.{conversation_history, percentage_used, total_tokens, used_tokens}, session_id, status, timestamp

Racket Code Search Support

  • crates/g3-core/src/code_search/searcher.rs
    • Racket parser [~45] - tree_sitter_racket::LANGUAGE
    • Extensions [~90] - .rkt, .rktl, .rktd → "racket"

Language-Specific Prompt Injection

Auto-detects languages and injects toolchain guidance.

  • crates/g3-cli/src/language_prompts.rs
    • LANGUAGE_PROMPTS - (lang_name, extensions, prompt_content)
    • AGENT_LANGUAGE_PROMPTS - (agent_name, lang_name, prompt_content)
    • detect_languages() - scans workspace
    • scan_directory_for_extensions() - recursive, depth 2, skips hidden/vendor
    • get_language_prompts_for_workspace(), get_agent_language_prompts_for_workspace()
  • crates/g3-cli/src/agent_mode.rs - appends agent-specific prompts
  • prompts/langs/ - language prompt files
  • To add language: Create prompts/langs/<lang>.md, add to LANGUAGE_PROMPTS
  • To add agent+lang: Create prompts/langs/<agent>.<lang>.md, add to AGENT_LANGUAGE_PROMPTS

MockProvider for Testing

  • crates/g3-providers/src/mock.rs
    • MockProvider [220..320] - response queue, request tracking
    • MockResponse [35..200] - configurable chunks and usage
    • scenarios module [410..480] - text_only_response(), multi_turn(), tool_then_response()
  • crates/g3-core/tests/mock_provider_integration_test.rs - integration tests
  • Usage: MockProvider::new().with_response(MockResponse::text("Hello!"))

G3 Status Message Formatting

  • crates/g3-cli/src/g3_status.rs
    • Status [12] - enum: Done, Failed, Error, Custom, Resolved, Insufficient, NoChanges
    • G3Status [44] - static methods for "g3:" prefixed messages
    • progress() [48], done() [72], failed() [81], thin_result() [236]

Prompt Cache Statistics

  • crates/g3-providers/src/lib.rs - Usage.cache_creation_tokens [6780], cache_read_tokens [6929]
  • crates/g3-providers/src/anthropic.rs - parses cache_creation_input_tokens, cache_read_input_tokens
  • crates/g3-providers/src/openai.rs - parses prompt_tokens_details.cached_tokens
  • crates/g3-core/src/lib.rs - CacheStats [3066]; Agent.cache_stats
  • crates/g3-core/src/stats.rs [189..230] - format_cache_stats() with hit rate metrics

Embedded Provider (Local LLM)

Local inference via llama-cpp-rs with Metal acceleration.

  • crates/g3-providers/src/embedded.rs
    • EmbeddedProvider [22..85] - session, model_name, max_tokens, temperature, context_length
    • new() [26..85] - tilde expansion, auto-downloads Qwen if missing
    • format_messages() [87..175] - converts to prompt string (Qwen/Mistral/Llama templates)
    • get_stop_sequences() [280..340] - model-specific stop tokens
    • stream() [560..780] - via spawn_blocking + mpsc

Chat Template Formats

Model Start Token End Token
Qwen <|im_start|>role\n <|im_end|>
GLM-4 [gMASK]<sop><|role|>\n <|endoftext|>
Mistral <s>[INST] [/INST]
Llama <<SYS>> <</SYS>>

Recommended GGUF Models

Model Size Use Case
GLM-4-9B-Q8_0 ~10GB Fast, capable
GLM-4-32B-Q6_K_L ~27GB Top tier coding/reasoning
Qwen3-4B-Q4_K_M ~2.3GB Small, rivals 72B

Download: huggingface-cli download <repo> --include "<file>" --local-dir ~/.g3/models/

Config:

[providers.embedded.glm4]
model_path = "~/.g3/models/THUDM_GLM-4-32B-0414-Q6_K_L.gguf"
model_type = "glm4"
context_length = 32768
max_tokens = 4096
gpu_layers = 99

Agent Skills System

Portable skill packages with SKILL.md + optional scripts per Agent Skills spec (agentskills.io).

  • crates/g3-core/src/skills/mod.rs [0..1501] - exports: Skill, discover_skills, generate_skills_prompt
  • crates/g3-core/src/skills/parser.rs [0..10750]
    • Skill [389] - name, description, metadata, body, path
    • Skill::parse() [1632] - parses SKILL.md with YAML frontmatter
    • validate_name() [4970] - 1-64 chars, lowercase+hyphens
  • crates/g3-core/src/skills/discovery.rs [0..12921]
    • discover_skills() [1266] - scans 5 locations: embedded → global → extra → workspace → repo
    • load_embedded_skills() [3263] - synthetic path <embedded:name>/SKILL.md
  • crates/g3-core/src/skills/embedded.rs [0..1674]
    • EmbeddedSkill [574] - name, skill_md
    • EMBEDDED_SKILLS [944] - static array (currently empty)
  • crates/g3-core/src/skills/prompt.rs [0..5628]
    • generate_skills_prompt() [397] - generates <available_skills> XML
  • crates/g3-config/src/lib.rs [180..200] - SkillsConfig (enabled, extra_paths)
  • crates/g3-cli/src/project_files.rs - discover_and_format_skills()

Skill Locations (priority: later overrides earlier):

  1. Embedded (compiled in)
  2. ~/.g3/skills/ (global)
  3. Config extra_paths
  4. .g3/skills/ (workspace)
  5. skills/ (repo root)

SKILL.md Format:

---
name: skill-name          # Required: 1-64 chars, lowercase + hyphens
description: What it does # Required: 1-1024 chars
license: Apache-2.0       # Optional
compatibility: Requires X # Optional
---
# Instructions...

Research Tool (First-Class)

Async web research via background scout agent.

  • crates/g3-core/src/pending_research.rs [0..18348]
    • ResearchStatus [682] - Pending/Complete/Failed
    • ResearchTask [1273] - task state
    • PendingResearchManager [2906] - thread-safe tracking with Arc
    • with_notifications() [3749] - broadcast channel for interactive mode
    • register() [5069], complete() [5480], fail() [6419], get() [7344], list_pending() [7806], take_completed() [8952]
  • crates/g3-core/src/tools/research.rs [0..17060]
    • CONTEXT_ERROR_PATTERNS [929] - detects context window exhaustion
    • execute_research() [1644] - spawns scout agent in background tokio task
    • execute_research_status() [7540] - check pending/completed
    • extract_report() [10694], strip_ansi_codes() [13148]
  • crates/g3-core/src/lib.rs
    • inject_completed_research() [31375] - injects results as user messages
    • enable_research_notifications() [33459] - for interactive mode
  • Tools: research (async, returns research_id), research_status (check pending)

Plan Mode

Structured task planning with cognitive forcing — requires happy/negative/boundary checks.

  • crates/g3-core/src/tools/plan.rs [0..49798]
    • PlanState [1044] - enum: Todo, Doing, Done, Blocked
    • Checks [2823] - happy, negative[], boundary[]
    • PlanItem [4021] - id, description, state, touches, checks, evidence, notes
    • Plan [6498] - plan_id, revision, approved_revision, items[]
    • EvidenceType [9578] - CodeLocation, TestReference, Unknown
    • VerificationStatus [10133] - Verified, Warning, Error, Skipped
    • parse_evidence() [12712] - parses file:line-line or file::test_name
    • verify_code_location() [14888] - checks file exists, lines in range
    • verify_test_reference() [16733] - checks test file, searches for fn
    • get_plan_path() [18655] - .g3/sessions/<id>/plan.g3.md
    • read_plan() [18818], write_plan() [19277] - YAML in markdown
    • plan_verify() [21978] - verifies evidence when complete; checks envelope existence
    • format_verification_results() [23395] - takes working_dir: Option<&Path> as third param
    • execute_plan_read() [25881], execute_plan_write() [27233], execute_plan_approve() [30651]
  • crates/g3-core/src/tool_definitions.rs [263..330] - plan_read, plan_write, plan_approve
  • crates/g3-core/src/prompts.rs [21..130] - SHARED_PLAN_SECTION
  • Tool names: plan_read, plan_write, plan_approve (underscores, not dots)
  • Evidence formats: src/foo.rs:42-118, src/foo.rs:42, tests/foo.rs::test_bar

Invariants System (Rulespec & Envelope)

Machine-readable invariants for Plan Mode verification. Rulespec read from analysis/rulespec.yaml (checked-in).

  • crates/g3-core/src/tools/invariants.rs [0..73975]
    • Claim [2024] - name + selector
    • PredicateRule [3009] - Contains, Equals, Exists, NotExists, GreaterThan, LessThan, MinLength, MaxLength, Matches
    • Predicate [5617] - claim, rule, value, source, notes
    • Rulespec [8734] - claims[] + predicates[]
    • ActionEnvelope [11203] - facts HashMap
    • Selector [12900] - XPath-like: foo.bar, foo[0], foo[*]
    • read_rulespec() [29472] - takes &Path (working_dir)
    • evaluate_rulespec() [32056] - evaluates against envelope

Write Envelope Tool

  • crates/g3-core/src/tools/envelope.rs [0..23347]
    • execute_write_envelope() [8764] - parses YAML facts, writes envelope.yaml, calls verify_envelope()
    • verify_envelope() [11705] - compiles rulespec on-the-fly, extracts facts, runs datalog, writes .dl + datalog_evaluation.txt (shadow mode)
  • crates/g3-core/src/tool_definitions.rs [266..282] - write_envelope tool definition
  • crates/g3-core/src/tool_dispatch.rs - write_envelope dispatch case
  • Workflow: write_envelope → verify_envelope() → datalog shadow, then plan_write(done) → plan_verify() → checks envelope exists

Datalog Invariant Verification

  • crates/g3-core/src/tools/datalog.rs [0..80172]
    • CompiledPredicate [1681] - id, claim_name, selector, rule, expected_value, source, notes
    • CompiledRulespec [2728] - plan_id, compiled_at_revision, predicates, claims
    • compile_rulespec() [3588] - validates selectors, builds claim lookup
    • Fact [6741] - claim_name, value
    • extract_facts() [7057] - uses Selector to navigate envelope YAML; fallback wraps in facts: if selector has facts. prefix
    • extract_values_recursive() [8478] - handles arrays/objects/scalars, adds __length facts
    • DatalogPredicateResult [10308], DatalogExecutionResult [10862]
    • execute_rules() [11627] - builds fact lookup, uses datafrog Iteration; when conditions delegate to evaluate_predicate_datalog()
    • evaluate_predicate_datalog() [14872] - handles all 9 PredicateRule types
    • escape_datalog_string() [23990], format_datalog_program() [24582] - Soufflé-style .dl output
    • format_datalog_results() [31136] - formats for shadow mode display
  • Relations: claim_value(claim, value), claim_length(claim, length), predicate_pass(id), predicate_fail(id)

Solon Agent (Rulespec Authoring)

  • agents/solon.md - interactive rulespec authoring agent prompt
  • crates/g3-cli/src/embedded_agents.rs [551] - 9 embedded agents: breaker, carmack, euler, fowler, hopper, huffman, lamport, scout, solon
  • Usage: g3 --agent solon

Structured Tool Call Messages

Native tool calls stored as structured MessageToolCall objects, not inline JSON text.

  • crates/g3-providers/src/lib.rs [0..17486]
    • MessageToolCall [2894] - id, name, input
    • Message [3014] - tool_calls: Vec<MessageToolCall>, tool_result_id: Option<String>
  • crates/g3-providers/src/anthropic.rs [0..74631]
    • convert_messages() [8642] - emits tool_use/tool_result blocks for structured tool calls
    • strip_orphaned_tool_use() [14737] - defense-in-depth: strips orphaned tool_use blocks with no matching tool_result
    • ToolResultContent [46268] - enum (Text | Blocks) for structured content
    • ToolResultBlock [46650] - enum (Image, Text) inside tool_result; images from read_image nested here, not as top-level blocks
  • crates/g3-core/src/lib.rs - ToolCall.id [2516] field from native providers
  • crates/g3-core/src/streaming_parser.rs [0..29244] - process_chunk() [10449] preserves tool call id
  • Gotcha: Images in tool result messages must be nested inside tool_result.content array, not as top-level Image blocks (Anthropic API rejects mixed top-level Image+ToolResult)

Tool Call Token Tracking

  • crates/g3-core/src/context_window.rs - estimate_message_tokens() [7695] accounts for tool_calls[].input
  • Token formula: content_tokens + per-tool (input_chars/3 * 1.1 + 20 overhead)
  • Gotcha: Without this, tool input JSON is invisible to tracker → compaction never triggers → API 400

Studio SDLC Pipeline

Orchestrates 7 agents in sequence for codebase maintenance.

  • crates/studio/src/sdlc.rs
    • PIPELINE_STAGES [28..62] - euler → breaker → hopper → fowler → carmack → lamport → huffman
    • Stage [18..26], StageStatus [65..80] - Pending, Running, Complete, Failed, Skipped
    • PipelineState [108..140] - run_id, stages[], commit_cursor, session_id
    • display_pipeline() [354..390] - box display with status icons
  • crates/studio/src/main.rs
    • cmd_sdlc_run() [540..655] - orchestrates pipeline, merges on completion
    • has_commits_on_branch() [715..728] - counts commits ahead of main
  • crates/studio/src/git.rs - merge_to_main() (hardcodes 'main')
  • State: .g3/sdlc/pipeline.json
  • CLI: studio sdlc run [-c N], studio sdlc status, studio sdlc reset

Terminal Width Responsive Output

Tool output responsive to terminal width — no line wrapping, 4-char right margin.

  • crates/g3-cli/src/terminal_width.rs
    • get_terminal_width() [21..28] - usable width (terminal - 4), min 40, default 80
    • clip_line() [33..44] - clips with "…", UTF-8 safe
    • compress_path() [53..96] - preserves filename, truncates dirs from left
    • compress_command() [101..103] - clips from right
    • available_width_after_prefix() [115..117]
  • crates/g3-cli/src/ui_writer_impl.rs
    • print_tool_output_header() [293..410] - uses compress_path/compress_command
    • update_tool_output_line() [407..445], print_tool_output_line() [447..454] - clip_line()
    • print_tool_compact() [475..635] - width-aware compact display

Plan Approval Gate (Non-Destructive + Baseline-Aware)

  • crates/g3-core/src/tools/plan.rs [973..983] - ApprovalGateResult enum: Allowed, Blocked { message }, NotGitRepo — no reverted_files field
  • crates/g3-core/src/tools/plan.rs [985..1003] - get_dirty_files() - returns HashSet<String> of dirty file paths from git status --porcelain
  • crates/g3-core/src/tools/plan.rs [1005..1098] - check_plan_approval_gate(session_id, working_dir, baseline_dirty) - warn-only, never reverts/deletes files, excludes baseline dirty files
  • crates/g3-core/src/lib.rs [170..171] - baseline_dirty_files: HashSet<String> field on Agent
  • crates/g3-core/src/lib.rs [1675..1686] - set_plan_mode(enabled, working_dir) - captures baseline on enable, clears on disable
  • Key invariant: The approval gate NEVER deletes or reverts files. It only warns.
  • Key invariant: Pre-existing dirty files (captured at plan mode start) are excluded from gate checks.

Context Window Calibration (Token Drift Fix)

  • crates/g3-core/src/context_window.rs [159..189] - update_usage_from_response() now calibrates used_tokens from API prompt_tokens (ground truth). When prompt_tokens > 0, snaps used_tokens to it. When 0, leaves unchanged (heuristic fallback).
  • crates/g3-core/src/context_window.rs [93..100] - No more 1% safety buffer. total_tokens = raw (was raw * 0.99).
  • crates/g3-core/src/context_window.rs [222..250] - estimate_message_tokens() now adds: +4 per-message overhead, +30 per tool_use block (was 20), +15 per tool_result message.
  • crates/g3-core/src/lib.rs [2232..2241] - ensure_context_capacity() called inside streaming loop for iteration > 1 (catches post-tool-execution growth).
  • Root cause: Heuristic token estimation drifted ~48% over 809 messages / 388 tool calls (136k estimated vs 201k actual). API prompt_tokens is ground truth.

Context Window Calibration (Token Drift Fix) - CORRECTED

  • crates/g3-core/src/context_window.rs [168..189] - update_usage_from_response() calibrates used_tokens from API prompt_tokens (ground truth). When prompt_tokens > 0, snaps used_tokens to it. When 0, leaves unchanged (heuristic fallback).
  • crates/g3-core/src/lib.rs [2316..2319] - Calibration call placed inline during streaming (when usage chunk arrives in chunk.usage), NOT after the streaming loop. Critical because text-only responses take an early return path that bypasses post-loop code.
  • crates/g3-core/src/lib.rs [2892..2898] - Post-loop code only handles fallback (no-usage) case now.
  • crates/g3-core/src/context_window.rs [87..93] - 1% safety buffer IS still in place (total_tokens * 0.99). Left as safety net between calibration points.
  • Root cause of display bug: (1) update_usage_from_response never calibrated used_tokens, only cumulative_tokens. (2) execute_single_task had mock usage with hardcoded prompt_tokens: 100. (3) Post-loop usage update was bypassed by early returns in text-only response paths.
  • Key streaming flow: For text-only responses (most common in interactive mode), chunk.finished triggers an early return Ok(self.finalize_streaming_turn(...)) that bypasses all post-loop code. Calibration MUST happen inline when chunk.usage arrives.
Reference in New Issue View Git Blame Copy Permalink